Decentralized personalized service system and method

Abstract

The invention provides a decentralized personalized service system and method, and the system is characterized in that an intelligent interaction module transmits user data uploaded by a user node to a data management module, and the data management module encrypts and stores the user data in a block chain; task related data is obtained from the data management module based on the task demand of the user node, and a task intelligent contract is issued according to the task related data, and a task AI model corresponding to the task demand is trained, constructed and stored by the artificial intelligence service module based on a preset federated learning mechanism; and after the service task is executed based on the task AI model corresponding to the task demand and the execution result of the service task is stored, the intelligent interaction module feeds back the execution result of the task intelligent contract to the user node. According to the method, user data leakage and abuse can be prevented, network attacks and security threats can be resisted, the accuracy and generalization ability of the model can be improved by fully utilizing distributed data resources, user intentions can be accurately understood, reliable services can be provided in time, and user satisfaction and loyalty can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of Internet information technology, and particularly to a decentralized personalized service system and method. Background Art

[0002] While Web3 technology, as a new generation of Internet technology, has been widely applied due to its emphasis on decentralization, user sovereignty over data, and blockchain-based value interaction, artificial intelligence technology has shown great advantages in data processing, analysis, and decision-making. It has become an inevitable trend to combine the two to provide an intelligent and secure data processing and application service system.

[0003] There are mainly the following two existing technical solutions for combining Web3 technology and artificial intelligence technology, but there are also certain application defects: One is to simply apply an artificial intelligence model to a decentralized application, but the model training data is mainly concentrated in a few nodes or collected in a centralized manner. There is a risk of privacy leakage of the data, and users cannot truly control their private data. Moreover, a large amount of data in the distributed network cannot be fully utilized, resulting in limited model training effects, insufficient model accuracy and generalization ability; The other is to perform simple data analysis on decentralized stored data, which can only achieve encrypted storage of data and basic access control, and there is insufficient value mining of the stored data. Complex operations such as intelligent classification and prediction of data cannot be performed, making it difficult to fully utilize the capabilities of artificial intelligence. The application functions are single, lacking intelligent interaction and personalized services, and the user experience is poor. Therefore, there is an urgent need to provide an innovative service system that can highly integrate artificial intelligence technology and Web3 technology to give full play to the advantages of both, achieve data privacy protection, efficient model training, intelligent interaction, and personalized services, and better meet the complex service requirements in the decentralized application scenario. Summary of the Invention

[0004] In order to solve the problems of the existing technology, it is necessary to provide a decentralized personalized service system and method for the above-mentioned technical problems.

[0005] In a first aspect, an embodiment of the present invention provides a decentralized personalized service system, characterized in that the personalized service system includes an intelligent interaction module, a data management module, and an artificial intelligence service module:

[0006] The intelligent interaction module is used to receive user data uploaded by each user node and transmit the user data to the data management module based on a data storage request; it is also used to obtain the task requirements of each user node, obtain task-related data corresponding to the task requirements from the data management module based on a data call request, and publish a task intelligent contract corresponding to the task requirements based on the task-related data, and in response to the completion of the execution of the task intelligent contract, feedback the execution result of the task intelligent contract to the user node; the task intelligent contract includes a model training task intelligent contract and a service task execution intelligent contract;

[0007] The data management module is used to encrypt and store the user data transmitted by the intelligent interaction module in blocks on the blockchain based on the data storage request; it is also used to transmit the task-related data corresponding to the task requirements to the intelligent interaction module based on the data call request;

[0008] The artificial intelligence service module is used to, in response to the execution of the model training task intelligent contract, train and construct a task AI model corresponding to the task requirements based on a preset federated learning mechanism, and store the task AI model on the blockchain; it is also used to, in response to the execution of the service task execution intelligent contract, execute a service task based on the task AI model corresponding to the task requirements, and store the obtained service task execution result on the blockchain.

[0009] In a second aspect, an embodiment of the present invention provides a decentralized personalized service method, and the method includes:

[0010] Receiving in real time user data uploaded by each user node, and encrypting and storing the user data in blocks on the blockchain;

[0011] Obtaining the task requirements of each user node, obtaining corresponding task-related data from the blockchain based on the task requirements, and publishing a task intelligent contract corresponding to the task requirements based on the task-related data; the task intelligent contract includes a model training task intelligent contract and a service task execution intelligent contract;

[0012] In response to the execution of the model training task intelligent contract, training and constructing a task AI model corresponding to the task requirements based on a preset federated learning mechanism, storing the task AI model on the blockchain, and feedbacking the execution result of the model training task intelligent contract to the user node;

[0013] In response to the execution of the service task execution intelligent contract, executing a service task based on the task AI model corresponding to the task requirements, storing the obtained service task execution result on the blockchain, and feedbacking the execution result of the service task execution intelligent contract to the user node.

[0014] Compared with the prior art, through the deep integration of Web3 technology and artificial intelligence technology, combined with decentralized data storage and multi-layer encryption technology, as well as an improved federated learning mechanism based on privacy protection and meta-learning, the present invention can not only prevent the leakage and abuse of user data during storage, training, and interaction, effectively resist various network attacks and security threats, and ensure the sustainable and stable operation of the system, but also make full use of the data resources of the distributed network to improve the accuracy and generalization ability of the constructed model, and can also accurately understand the user's behavior intention based on multi-modal input and provide reliable service feedback in a timely manner, making the interaction more natural and convenient, thereby realizing more intelligent, secure, and efficient data processing and application services, and effectively improving user satisfaction and loyalty. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a schematic structural diagram of a decentralized personalized service system in an embodiment of the present invention;

[0016] Figure 2 is a schematic diagram of the application scenario of the data management module in an embodiment of the present invention;

[0017] Figure 3 is a schematic diagram of the federated learning mechanism of the intelligent contract for performing model training tasks in an embodiment of the present invention;

[0018] Figure 4 is another schematic structural diagram of a decentralized personalized service system in an embodiment of the present invention;

[0019] Figure 5 is a schematic diagram of the application scenario of the incentive management module in an embodiment of the present invention;

[0020] Figure 6 is a schematic diagram of the application scenario of the security protection module in an embodiment of the present invention;

[0021] Figure 7 is a schematic flowchart of a decentralized personalized service method in an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] In order to make the purpose, technical solutions, and beneficial effects of the present application clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the following described embodiments are part of the embodiments of the present invention and are only used to illustrate the present invention, but not to limit the scope of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0023] In some embodiments, such as Figure 1As shown, a decentralized personalized service system is provided. The system includes an intelligent interaction module 1, a data management module 2, and an artificial intelligence service module 3;

[0024] The intelligent interaction module 1 is used to receive user data uploaded by each user node and transmit the user data to the data management module 2 based on a data storage request; it is also used to obtain the task requirements of each user node, obtain task-related data corresponding to the task requirements from the data management module 2 based on a data call request, and publish a task smart contract corresponding to the task requirements based on the task-related data. In addition, in response to the completion of the execution of the task smart contract, the execution result of the task smart contract is fed back to the user node; the task smart contract includes a model training task smart contract and a service task execution smart contract.

[0025] The intelligent interaction module 1 can be understood as a management module for the intelligent interaction function between the user and the personalized service system. The user data it receives can be understood as data collected or gathered by each user himself / herself and transmitted to the personalized service system for storage and use through the corresponding terminal device. It can include text data, voice data, pictures, or other format data related to the personalized services that the user is concerned about. It should be noted that when each user node uploads user data, it can specify corresponding usage permissions and data classification and other information. After the intelligent interaction module 1 receives the user data uploaded by the user node based on the corresponding intelligent interaction interface, it generates a corresponding data storage request based on the user data and the corresponding usage permissions and data classification and other information and sends it to the data management module 2. The data management module 2 stores the user data on the chain according to the information carried in the data storage request.

[0026] In addition to receiving user-uploaded storage data, the intelligent interaction module 1 can also perform intent analysis on the user input data to accurately create and execute service tasks that meet the user's intent, and promptly feed back the corresponding service task execution results to the user, realizing efficient intelligent interaction between the user and the system, providing personalized recommendations and feedback based on the user's behavior and preferences, and enhancing the user experience. In some embodiments, the intelligent interaction module 1 includes a requirement analysis unit and a requirement confirmation unit;

[0027] The requirement analysis unit is used to perform intent analysis on the task input data obtained based on the intelligent interaction interface to obtain task requirements corresponding to the task input data; among them, both the task input data and the user data can be obtained based on the intelligent interaction interface, but they can be uploaded through different channel entrances. For example, the user data can be directly uploaded based on the channel entrance for dedicated data storage, and the task input data can be uploaded based on the channel entrance for dedicated user task publishing. No specific limitation is made here.

[0028] To ensure the accuracy and comprehensiveness of the analysis of the user's task - publishing intention, the embodiments of the present invention support the user to provide data inputs in multiple formats; in some embodiments, the task input data includes at least one of text data, voice data, and image data; the requirement analysis unit includes a text data analysis unit, a voice data analysis unit, an image data analysis unit, and a comprehensive analysis unit;

[0029] The text data analysis unit is used to perform semantic analysis on the text data based on natural language processing technology to generate text data requirements; among them, the natural language processing (Natural Language Processing, NLP) technology can be selected according to actual application requirements. For example, a pre - trained model based on the Transformer architecture can be used for lexical, syntactic, and semantic analysis to understand the user intention shown in the text data, that is, to obtain the required text data requirements.

[0030] The voice data analysis unit is used to perform recognition and conversion on the voice data based on voice recognition technology to obtain voice text data, and perform semantic analysis on the voice text data based on the natural language processing technology to generate voice data requirements; among them, the voice recognition technology can also be selected according to actual application requirements. For example, the DeepSpeech model can be used to convert the received voice data into text data, and then the natural language processing technology used in the above - mentioned text data analysis is used for analysis to obtain the required voice data requirements.

[0031] The image data analysis unit is used to perform feature extraction and analysis on the image data based on computer vision technology to generate image data requirements; among them, the computer vision technology can also be selected according to actual application requirements. For example, a convolutional neural network or a graph convolutional neural network, etc. can be used to perform feature extraction and analysis on the image data to obtain the image data requirements.

[0032] The comprehensive analysis unit is used to perform comprehensive requirement analysis based on the text data requirements, the voice data requirements, and / or the image data requirements to generate the task requirements; among them, the comprehensive requirement analysis can be understood as comparing the analysis results of various types of data involved in the task input data, merging the repeated results, and merging or screening the different results according to preset rules to ensure the comprehensiveness and rationality of the finally obtained task requirements.

[0033] The requirement confirmation unit is used to feedback the task requirement to the user node corresponding to the task input data based on the intelligent interaction interface, so as to confirm or adjust the task requirement; after the task requirement is obtained through the above method, in order to improve the accuracy of understanding the task requirement (user intention) as much as possible, the task requirement obtained by the system intelligent analysis is also feedback to the user through the intelligent interaction interface for confirmation, and the currently generated task requirement is modified, adjusted or supplemented in a timely manner according to the information feedback by the user, so as to ensure the reliability and accuracy of subsequent user task execution.

[0034] After determining the user's task requirement through the above method steps, the user task type (training model task or service execution task) can be determined based on this task requirement, and then the required data resources can be determined according to the specific user task type, and the corresponding data call request can be generated and sent to the data management module 2, and receive the task-related data (which may include task analysis data and / or task AI model) corresponding to the data call request from the data management module 2, and release the task intelligent contract (model training task intelligent contract and service task execution intelligent contract) corresponding to the task requirement based on the task-related data and the preset intelligent contract release rule, and after the corresponding task intelligent contract is executed, the execution result of the task intelligent contract is feedback to the user node; it should be noted that since the data managed by the data management module 2, whether it is user data (task analysis data) or task AI model, etc., can be understood as data already stored on the blockchain, when the data management module 2 feedbacks the task-related data according to the data resource type in the data call request, it can directly feedback the block address information corresponding to each data without directly transmitting the data itself, reducing the communication bandwidth consumption and avoiding the security risk brought by data transmission at the same time; in addition, the operation mechanism of the task intelligent contract released by the intelligent interaction module 1 can be implemented according to the actual application requirements in combination with the operation mechanism of the existing intelligent contract on the blockchain, which is not specifically limited here.

[0035] The data management module 2 is used to encrypt and store the user data transmitted by the intelligent interaction module 1 in blocks on the blockchain based on the data storage request; it is also used to transmit the task-related data corresponding to the task requirement to the intelligent interaction module 1 based on the data call request.

[0036] Encrypted block storage can be understood as the process of first encrypting the user data to be stored using multi-layer encryption technology and then dividing it into fixed-size data blocks for storage on the blockchain; in order to achieve decentralized data storage while ensuring user data privacy and supporting efficient data retrieval, in some embodiments, as Figure 2 shown, the data management module 2 includes an encryption block unit and a data storage unit;

[0037] The encryption block unit is used to obtain encrypted user data based on the encryption process performed on the user data of each user node, and obtain multiple encrypted data blocks based on the segmentation process performed on each encrypted user data. Among them, the encryption process can be implemented using a multi-layer encryption algorithm. For example, the data is first encrypted using a symmetric encryption algorithm (AES-256), and then the symmetric encryption key is encrypted using an asymmetric encryption algorithm (RSA-2048). In the embodiments of the present invention, the sizes of all encrypted data blocks are the same, that is, the encrypted user data is segmented into data blocks of a fixed size (such as each data block is 1MB), and a unique hash value is assigned to each encrypted data block as the data block identifier.

[0038] The data storage unit is used to store each encrypted data block in the blockchain based on the distributed hash table technology, and store the metadata corresponding to each encrypted data block in the blockchain based on the content-addressable storage technology. Among them, the distributed hash table (DHT) technology is a distributed key-value storage system, which stores each encrypted data block and the corresponding hash value as key-value pairs on different block nodes of the blockchain, and the data consistency and integrity are ensured among different block nodes through the consensus mechanism of the blockchain.

[0039] To facilitate the rapid retrieval of each encrypted data block, metadata corresponding to each encrypted data block is set, including data source (such as user device ID), data type (such as image data, text data, voice data, and transaction data, etc.), timestamp, data block hash value, and access rights (such as whether the data is allowed to be shared, the shared data range, and the shared user range, etc., so that each user can truly and effectively control private data and reasonably share other users' data). Similarly, to ensure the security of the metadata, after it is encrypted, it is stored on a specific blockchain node based on the content-addressable storage (CAS) technology, and the encrypted data block can be quickly located according to the data block hash value in the metadata.

[0040] After the data management module 2 stores the user data on the blockchain through the above method, it also transmits the task-related data corresponding to the task requirements to the intelligent interaction module 1 based on the data call request of the intelligent interaction module 1. For example, when it is necessary to query the transaction data of a certain user at a specific time, after the system retrieves the matching metadata in the CAS according to the metadata information such as transaction data type and timestamp, the corresponding encrypted data block can be located based on the data block hash value in the metadata, and the block address corresponding to the required encrypted data block is fed back to the intelligent interaction module 1 for it to execute the corresponding task smart contract.

[0041] The artificial intelligence service module 3 is used to, in response to the execution of the model training task smart contract, train and construct a task AI model corresponding to the task requirements based on a preset federated learning mechanism, and store the task AI model in the blockchain; it is also used to, in response to the execution of the service task execution smart contract, execute the service task based on the task AI model corresponding to the task requirements, and store the obtained service task execution result in the blockchain.

[0042] The artificial intelligence service module 3 can be understood as a functional module for executing task smart contracts, which can not only execute the training task of the task AI model, but also execute the service task of analyzing based on the task AI model obtained by pre-training. The above-mentioned model training task smart contract and service task execution smart contract can be set according to actual application requirements. For example, the model training task smart contract can include contract basic information (such as contract name, contract version, and contract participants), participant permission allocation (data provider, model trainer, model user, etc.), model training requirement information (model training goal: such as the prediction accuracy rate reaches more than 90%, training data type: such as user behavior data and transaction data, training model parameters: number of neural network layers and number of neurons, etc., required computing resources and required storage resources, etc.), and model training management (such as including model training process, event trigger status mechanism, and reward mechanism), etc.; the service task execution smart contract can include contract basic information and service requirement information (required analysis data type and required task AI model type, etc.).

[0043] In some embodiments, the artificial intelligence service module 3 includes a model training unit and a model execution unit;

[0044] The model training unit is used to generate a model training task based on the model training requirement information in the model training task smart contract, perform task splitting training on the model training task based on the preset federated learning mechanism, obtain multiple encrypted subtask model parameters, and perform weighted aggregation on all encrypted subtask model parameters based on the first aggregation smart contract to obtain the task AI model; the model training requirement information includes model training goal, training data type, and training model parameters.

[0045] The preset federated learning mechanism in the embodiments of the present invention can be understood as a distributed federated learning mechanism for task splitting training as Figure 3 shown; in some embodiments, the model training unit includes a task splitting subunit, a subtask training subunit, a model aggregation subunit, and a model evaluation subunit;

[0046] The task splitting sub-unit is used to send the model training task to the main aggregation node, and the main aggregation node splits the model training task into multiple sub-training tasks based on the preset federated learning mechanism and distributes them to different secondary aggregation nodes; among them, the main aggregation node can be understood as a model training participating node used to split model training into multiple sub-tasks and reliably aggregate the models corresponding to the multiple sub-tasks; the corresponding secondary aggregation node can be understood as a lower-level model training participating node subordinate to the main aggregation node, which is used to assist multiple underlying training nodes to complete the sub-training tasks sent by the main aggregation node. It should be noted that the splitting method of the model training task varies according to the specific task situation, and the main aggregation node, secondary aggregation node and training node can be determined by the scoring mechanism in actual applications, which is not specifically limited here.

[0047] The sub-task training sub-unit is used to distribute the corresponding sub-training tasks to the corresponding multiple training nodes by each secondary aggregation node, and perform weighted aggregation on the encrypted local model parameters obtained by each training node based on the preset training smart contract to obtain the corresponding encrypted sub-task model parameters and upload them to the main aggregation node; among them, the preset training smart contract may include the usage method of training data, the number of training iterations, the loss function, the model training strategy, etc.

[0048] In order to ensure that each training node has sufficient training data to meet the generalization of local model training, in some embodiments, the encrypted training data used by each training node to perform model training includes encrypted data obtained by interacting with the data management module and / or locally stored encrypted data; that is, each training node can extract the locally stored encrypted data according to the required data type involved in the sub-training task and obtain the same type of encrypted data authorized by other users on the blockchain through the data management model.

[0049] In order to avoid the risk of user data leakage during the training process, encrypted data can be directly used for model training. In some embodiments, the preset training smart contract includes performing model training using encrypted training data based on homomorphic encryption technology; that is, during the training process, homomorphic encryption technology is used to enable each training node to directly perform calculations on the encrypted data. For example, when calculating the gradient update, the training node does not need to decrypt the data. After the calculation is completed, the training node sends the updated encrypted local model parameters to the corresponding secondary aggregation node, and then the secondary aggregation node aggregates all the received encrypted local model parameters to update the encrypted sub-task model parameters, and distributes the new encrypted sub-task model parameters to the training nodes again, and iterates continuously until the encrypted sub-task model converges, and then the secondary aggregation node uploads it to the main aggregation node. The method of performing model training using encrypted training data based on homomorphic encryption technology can effectively ensure the security and privacy of user data while expanding the local training data set and improving the effect of local model training.

[0050] To ensure that there is no significant deviation in the data distribution of each training node during the subtask training process, so as to avoid the subtask model being biased towards the data of some training nodes and reducing the problem of the model's generalization ability, it is also possible to continuously monitor the distribution of the local model performance indicators of each training node and adjust the training strategy or data processing method in a timely manner. In some embodiments, the preset training smart contract includes adjusting the model training strategy based on the distribution difference between the distribution of the local model performance indicators of each training node and the subtask model performance indicators of the corresponding sub-aggregation node. For example, statistical methods (such as KL divergence) can be used to regularly check the distribution of the local model performance indicators of the training nodes in different rounds, calculate the difference between it and the subtask model performance indicators of the upper-level aggregation node, and determine whether the corresponding difference exceeds the preset distribution difference threshold. When it is determined that the difference exceeds, it is determined that the distribution of the local model performance indicators of this training node has deviated from the subtask model performance indicators, and the subsequent model training strategy needs to be adjusted accordingly.

[0051] To avoid the privacy leakage of each training node, differential privacy can also be introduced to add noise to the local model parameters of each training node and then transmit them to the corresponding sub-aggregation node; in some embodiments, the encrypted local model parameters of each training node are obtained by adding noise to the local model parameters obtained by performing model training on each training node based on the preset local differential privacy mechanism.

[0052] In order to more reasonably manage the privacy budget while using differential privacy to prevent the privacy leakage of each training node and avoid the overall privacy leakage risk, different privacy budgets can also be allocated according to the situations of different training nodes and task stages; in some embodiments, the privacy budget in the preset local differential privacy mechanism is set based on the iterative training progress, initial privacy budget, and data sensitivity of each training node; the privacy budget is inversely proportional to the iterative training progress; the privacy budget is directly proportional to the initial privacy budget and the data sensitivity. For example, a larger privacy budget is used in the initial stage, and as the number of training rounds increases, the privacy budget is gradually reduced. At the same time, considering the sensitivity of the data of different participants, more privacy budget is allocated to sensitive data; that is, the privacy budget is expressed as:

[0053] e r = e0 * (1 - r / r max ) * f c

[0054] In the formula, e0 represents the initial privacy budget; e r represents the privacy budget of the r-th iteration; r represents the current iteration number; r max represents the maximum number of iterations; f cRepresents a data sensitivity factor, which expresses the privacy protection requirements of the data. The greater the factor, the greater the privacy protection requirements, indicating that the more sensitive the data is.

[0055] The model aggregation subunit is used for the master aggregation node to perform weighted aggregation on the encrypted sub-task model parameters of each secondary aggregation node based on the first aggregation smart contract to obtain global model parameters. Among them, the first aggregation smart contract can be understood as a smart contract generated based on a preset model aggregation strategy, and aggregation methods such as weighted average or federated average can be used to obtain global model parameters.

[0056] The model evaluation subunit is used to evaluate the global model parameters based on a preset model evaluation smart contract. When the global model parameters do not meet the model training target, the model training task is updated based on the global model parameters, and iterative training continues based on the updated model training task until the obtained global model parameters meet the model training target. Then, the task AI model is obtained based on the corresponding global model parameters. Among them, the preset model evaluation smart contract can be understood as a smart contract that uses various evaluation indicators such as cross-validation and confusion matrix to evaluate the model performance from different perspectives to ensure the effectiveness and generalization ability of the global model. After aggregating the encrypted sub-task model parameters of each secondary aggregation node to obtain global model parameters, it is also necessary to evaluate and feedback the obtained aggregation result. If the evaluation result is not ideal or does not meet the requirement standard of the preset model evaluation smart contract, the specification parameters need to be adjusted and the parameters are feedback to the secondary aggregation nodes. Each secondary aggregation node then adjusts the sub-task specification parameters and feedbacks the corresponding parameters to each training node. Each training node performs iterative training according to the adjusted parameters and starts a new round of training until it meets the evaluation requirements of the preset model evaluation smart contract.

[0057] To further enhance the privacy protection intensity, the Secure Multi-Party Computation (SMPC) protocol can also be used to enable multiple participating parties to jointly complete more complex computing tasks without exposing their respective data, so as to ensure data privacy and computing security. In some embodiments, weighted aggregation is performed on the encrypted local model parameters of each training node based on secure multi-party computing technology; and / or, weighted aggregation is performed on the encrypted sub-task model parameters of each secondary aggregation node based on secure multi-party computing technology. For example, the secure multi-party computing protocol is used to implement operations such as encrypted multiplication and encrypted summation of model parameters to improve the security of data during the computing process.

[0058] To further optimize the model aggregation effect, while considering improving the weighting strategy, meta-learning ideas are also introduced to enhance the efficiency and reliability of model aggregation. In some embodiments, when performing weighted aggregation on the encrypted local model parameters of each training node, the aggregation weights of each training node are obtained by combining preset weight allocation factors with a meta-learning architecture; the preset weight allocation factors include the data quality, data distribution, and model performance of each training node. Among them, the meta-learning architecture can adopt the MAML (Model-Agnostic Meta-Learning) idea to learn how to better aggregate the encrypted local model parameters of different training nodes to obtain encrypted sub-task model parameters to adapt to the changes in the data quality, data distribution, and model performance of different training nodes. In practical applications, the encrypted local model parameters θ k of each training node, the data quality q k (which can use relevant evaluation indicators of the training effect of the existing model), the data distribution d k (such as the mean square error of the data), and the model performance p k (such as the accuracy, precision, and F1 score of the model, etc.) are used as model inputs, and the aggregation weight w k of each training node is used as the model output to construct a meta-model Assuming the meta-model parameters are and the encrypted sub-task model parameters are θ, then there is The aggregation of the encrypted local model parameters of each training node using the aggregation weights generated by the meta-model to obtain the encrypted sub-task model parameters can be expressed as K represents the total number of training nodes participating in the training. It should be noted that the training process of obtaining the aggregation weights of each training node by adopting the MAML meta-learning architecture based on the preset weight allocation factors of each training node can be implemented by referring to the existing MAML meta-learning training technology. Updating the parameters of the aggregation strategy during aggregation can enable the encrypted sub-task model parameters to adapt to the new data distribution and the addition or withdrawal of training nodes more quickly.

[0059] Considering that in practical applications, training nodes upload encrypted local model parameters and secondary aggregation nodes need to upload encrypted sub-task model parameters, and a large amount of data will reduce communication efficiency, thus affecting model aggregation efficiency. To improve the communication efficiency between nodes, in some embodiments, each training node uploads the encrypted local model parameters compressed based on sparse quantization technology to the corresponding secondary aggregation node; and / or, each secondary aggregation node uploads the encrypted sub-task model parameters compressed based on sparse quantization technology to the master aggregation node. That is, when each training node and / or each secondary aggregation node uploads the corresponding model parameters, they first compress their model parameters based on sparse quantization technology and only upload important model parameters; for example, using stochastic gradient descent sparsification, only retaining the elements in the gradient whose absolute value is greater than a preset threshold and setting other elements to zero to achieve sparsification, and then only transmitting the non-zero elements in the gradient vector; it is also possible to sample and map the floating-point values of the gradient to a fixed number of bits, converting the parameters from high precision to low precision representation to reduce the amount of data.

[0060] At the same time, considering that the existing federated learning mechanisms all adopt synchronous communication (aggregation is carried out only after all training nodes have completed local training), to further improve the communication effect, asynchronous communication can be further adopted, allowing training nodes to send updates immediately after completing training or after secondary aggregation nodes complete aggregation. The secondary aggregation node / master aggregation node adjusts the aggregated model at any time according to the received updates, reducing the waiting time and improving the overall efficiency of the system. In some embodiments, each training node uploads the encrypted local model parameters to the corresponding secondary aggregation node based on a preset asynchronous communication mechanism, and the secondary aggregation node continuously updates the encrypted sub-task model parameters based on a preset intelligent contract for sub-task model parameter updates; and / or, each secondary aggregation node uploads the encrypted sub-task model parameters to the master aggregation node based on a preset asynchronous communication mechanism, and the master aggregation node continuously updates the global model parameters based on a global model parameter update intelligent contract; among them, the global model parameter update intelligent contract can be determined according to actual application requirements, such as setting a fixed aggregation update period or setting the number of model update groups that trigger aggregation updates; that is, in practical applications, a mechanism of using a thread to maintain a message queue can be used to support the asynchronous transmission of corresponding model parameters by training nodes / secondary aggregation nodes, and the secondary aggregation node / master aggregation node at the other end continuously receives updates and immediately updates the aggregated model after determining that the update conditions of the global model parameter update intelligent contract are met.

[0061] It should be noted that in the two-layer aggregation federated learning architecture provided in the embodiments of the present invention, there is usually interaction between secondary aggregation nodes, and the main purpose is to more effectively perform local model aggregation and information fusion to assist the master aggregation node in completing the update of the global model. The general interaction information between them includes model parameter-related information, model evaluation and status information, and data statistical information.

[0062] The information related to model parameters includes: 1) Partially aggregated model weights. After the sub-aggregation nodes collect the model weights of the training nodes within their respective jurisdictions, they will perform preliminary aggregation. These partially aggregated model weights need to be exchanged among the sub-aggregation nodes to further integrate the model information of different regions. For example, after the sub-aggregation nodes in different regions preliminarily aggregate the model weights of the training nodes within their respective jurisdictions, they exchange these partially aggregated model weights to provide more comprehensive information for the total aggregation; 2) Intermediate gradient information. The sub-aggregation nodes may calculate intermediate gradient information and exchange it among each other. These intermediate gradient information helps the sub-aggregation nodes understand the change trends of model parameters in other regions, so as to adjust the local aggregation strategy more accurately. For example, in the federated learning task of image recognition, the sub-aggregation nodes can coordinate the learning of different types of image features by exchanging intermediate gradient information.

[0063] The model evaluation and status information mainly includes: 1) Local model evaluation metrics sum. The sub-aggregation nodes will calculate the evaluation metrics of the models within their respective jurisdictions, such as accuracy, recall rate, loss value, etc. The interaction of these local model evaluation metrics among the sub-aggregation nodes can help each sub-aggregation node understand the model performance in other regions, so as to adjust the local model training strategy in a timely manner; 2) Training status information: including whether the local training converges, the progress of training, whether there are any abnormalities, etc. Sharing this information among the sub-aggregation nodes can better coordinate the overall training process. If a sub-aggregation node finds that the local training does not converge, it can interact with other sub-aggregation nodes to check the cause of the problem and adjust the hyperparameters of the local training if necessary.

[0064] The data statistical information mainly includes: 1) Local data feature distribution. The sub-aggregation nodes know the feature distribution of the data within their respective regions, such as the mean, variance, and proportion of different feature values of the data. Exchanging this information can let the sub-aggregation nodes understand the distribution differences of the overall data, which helps to consider the characteristics of data in different regions when aggregating the model and improve the generalization ability of the model; 2) Data volume and sample diversity: The amount of client data and the sample diversity of the data responsible for each sub-aggregation node may be different. Exchanging this information can help the sub-aggregation nodes reasonably allocate weights when aggregating the model according to the size of the data volume and the degree of sample diversity, avoiding model bias caused by uneven data volume or single sample.

[0065] The model execution unit is configured to generate a corresponding service task based on the service requirement information in the smart contract for the service task, obtain service-related data corresponding to the service task through interaction with the data management module, and execute the service task based on the service-related data and a preset service execution smart contract to obtain a service task execution result; the service-related data includes service-related user data and a corresponding task AI model. Among them, the service requirement information in the service task execution smart contract may include, as described above, the types of analysis data required for the service task and the types of task AI models to be used, etc. When executing the service task execution smart contract, a corresponding data request message will be generated according to the service requirement information and sent to the data management module 2 to obtain the block storage address information of the required service-related data. After obtaining the corresponding data information based on the corresponding block storage address information, the service-related user data will be input into the corresponding task AI model according to the preset service execution smart contract to execute the corresponding service task to obtain the required service task execution result. For example, if the current service task is to query the price trend of a certain commodity, then the commodity transaction data, relevant market data, and the corresponding price prediction model can be obtained based on the data management module 2 for prediction analysis, and the corresponding prediction result can be fed back to the user through the intelligent interaction interface. If the user triggers the service task by text input, the result will be displayed in text form. If the user triggers the service task by voice input, the result will be converted into voice and played to the user through voice synthesis technology (such as the Tacotron model). In addition, personalized content recommendations and operation suggestions can be provided for the user through a recommendation algorithm based on the user's historical behavior data and preference information.

[0066] In the embodiments of the present invention, the intelligent interaction module transmits the user data uploaded by the user node to the data management module, encrypts and stores it in chunks on the blockchain, and based on the task requirements obtained from the user node, obtains task-related data from the data management module and publishes a task smart contract accordingly. The artificial intelligence service module trains, constructs, and stores a task AI model corresponding to the task requirements based on a preset federated learning mechanism, and based on the task AI model corresponding to the task requirements, executes the service task and stores the execution result of the service task. Then, the intelligent interaction module feeds back the execution result of the task smart contract to the user node in a decentralized personalized service system. By deeply integrating Web3 technology and artificial intelligence technology, combining decentralized data storage and multi-layer encryption technology, as well as an improved federated learning mechanism based on privacy protection and meta-learning, the application advantages of Web3 technology and artificial intelligence technology are fully utilized to meet the complex requirements in the decentralized application scenario. It can not only prevent the leakage and abuse of user data during storage, training, and interaction, effectively resist various network attacks and security threats, and ensure the sustainable and stable operation of the system, but also make full use of the data resources of the distributed network, improve the accuracy and generalization ability of the constructed model, and accurately understand the user's behavior intention based on multi-modal input and provide reliable service feedback in a timely manner, making the interaction more natural and convenient. Furthermore, it can achieve more intelligent, secure, and efficient data processing and application services, effectively improving user satisfaction and loyalty.

[0067] In addition, to ensure the sustainable and stable operation of the personalized service system, reward and punishment measures are also implemented for the system member nodes through a preset reward and punishment smart contract, and a security audit mechanism is established to regularly check the blockchain nodes, smart contracts, data storage, and transmission, etc., to promote the healthy development of the system. In some embodiments, as Figure 4 shown, the personalized service system further includes an incentive management module and a security protection module:

[0068] The incentive management module 4 is used to evaluate the contributions of the behaviors of each member node in real time, and based on the obtained behavior contribution evaluation results and a preset reward and punishment smart contract, execute an incentive mechanism, generate corresponding reward and punishment measures, and store them in the blockchain; the member node behaviors include data sharing, model training participation, and system maintenance. In practical applications, economic incentives, point systems, resource sharing, etc. can be used to provide incentives for the participants in the training task model, and different incentive values are given according to different contribution situations to encourage them to participate in federated learning more actively and improve data quality and participation. For example, a certain economic reward (automatically distributed based on the smart contract) is given according to the contributions of the participating parties (such as data volume, data quality, model performance improvement, etc.), or a point system is established among the participating parties, and the participating parties with higher points can preferentially obtain the right to use the global model or obtain other resources. At the same time, factors such as the security maintenance effect of the member nodes (such as the value of discovering and fixing security vulnerabilities) can also be considered to determine the reward amount, while malicious behaviors (such as providing false data, attacking the system, etc.) are punished, and the corresponding punishment measures include deducting cryptocurrency, restricting the permissions of nodes or users in the system (such as restricting data access permissions, prohibiting participation in model training), etc. All reward and punishment operations are recorded on the blockchain to ensure fairness and justice, as Figure 5 shown.

[0069] The security protection module 5 is used to regularly conduct security audits and vulnerability checks on the member nodes, smart contracts, data storage, and data transmission in the personalized service system based on a preset security audit mechanism, and store the obtained audit and inspection results in the blockchain, as Figure 6 shown. Among them, the preset security audit mechanism can be designed according to actual application requirements. For example, professional code audit tools and manual reviews are combined to check whether there are vulnerabilities in the smart contract (such as re-entrancy vulnerabilities, timestamp dependence vulnerabilities, etc.), and the discovered vulnerabilities are repaired in a timely manner.

[0070] The interaction relationships among the various modules in the personalized service system provided by the embodiments of the present invention are as Figure 4As shown, when the user submits task input data in the system through the intelligent interaction module 1, the intelligent interaction module 1 first accurately obtains the user's task requirements through intelligent semantic analysis, and then calls the corresponding data in the data management module 2 (decentralized data network module) to generate a corresponding task intelligent contract. When it is determined that the task requirement is a model training task, using the federated learning mechanism, the task is distributed to the main aggregation node, and then the main aggregation node further decomposes the task and distributes it to multiple secondary aggregation nodes and activates the active "worker" user nodes on the blockchain to use their idle resources to perform the corresponding task AI model training. During the process, homomorphic encryption technology and multi-party computing aggregation technology are used to improve data processing efficiency and ensure security, and continuous iteration and optimization are carried out until the final training result is completed and intelligent delivery is performed. And when it is determined that the task requirement is a service task, the corresponding analysis and usage data and task AI model in the data management module 2 (decentralized data network module) are called, and then the corresponding intelligent contract is executed to generate a corresponding service execution result and perform intelligent delivery. At the same time, during the entire system operation process, log files such as operation interaction records are collected in real time and corresponding audit results are generated (if there are abnormal situations, start the intelligent contract for warning and automatic termination). In addition, fully considering the attributes of the web3 system, this system integrates incentive and punishment mechanisms, uses the built-in scoring mechanism module to evaluate the work conditions of all on-chain users participating in the relevant work of this system, and conducts the operation process of rewarding or punishing through intelligent contracts, thereby promoting the healthy development of this system. It should be noted that in the personalized service system in the above embodiments, a variety of encryption technologies are comprehensively used to ensure the privacy of user data. During the data storage and transmission process, homomorphic encryption is used to ensure data calculation privacy, and at the same time, zero-knowledge proof (ZKP) is also used to verify the attributes of data and the permissions of users. For example, when a user accesses restricted data, it can prove that it has access rights without revealing the specific data content.

[0071] To facilitate the understanding of the operation mechanism of the decentralized personalized service system provided by the present invention, the embodiments of the present invention also provide a simulation experiment taking the personalized commodity recommendation service in the e-commerce application scenario as an example:

[0072] 1) Environment construction. In a simulated decentralized e-commerce application scenario, a network composed of 100 nodes is constructed. Each node is configured with sufficient computing resources (such as CPU, GPU) and storage resources, runs the blockchain node software based on Ethereum, deploys the decentralized personalized service of the present invention, and configures relevant encryption algorithm parameters, such as the key length of the AES-256 algorithm, the public key and private key generation rules of RSA-2048, etc.

[0073] 2) Data storage and model training. An e-commerce platform can upload data such as the user's browsing history and purchase records on the platform to the system through the intelligent interaction module 1. The data is encrypted and stored in the data management module 2. When the e-commerce platform or other system user nodes issue a model training task for product recommendation with a recommendation accuracy rate of over 80%, the system will create a corresponding intelligent contract for the model training task. When the intelligent contract for the model training task is executed, the artificial intelligence service module 3 divides the corresponding model training task into multiple sub-product recommendation model training tasks based on the main aggregation node within the federated learning mechanism and distributes them to the participating nodes via multiple secondary aggregation nodes. After each participating node obtains the corresponding task information and initial product recommendation training model parameters, it can use the encrypted data stored locally or the shared encrypted data stored on the blockchain for training. During the training process, each training node uses homomorphic encryption technology to calculate the model parameter updates and sends the encrypted updates to the corresponding secondary aggregation node. The secondary aggregation node uses the Secure Multiparty Computation (SMPC) technology to aggregate and update the model parameters of each subordinate training node to obtain the corresponding encrypted sub-task model parameters, and then uploads them to the main aggregation node. The main aggregation node uses the SMPC technology to aggregate the encrypted sub-task model parameters to obtain the product recommendation AI model. After multiple iterations, when the model accuracy rate reaches 82%, it can be directly deployed on the blockchain for use.

[0074] 3) Intelligent interaction service. When a user inputs "Recommend summer clothes suitable for me" by voice through the intelligent interaction module 1, the speech recognition technology in the intelligent interaction module 1 converts the voice into text. Then, when it uses NLP technology to analyze and determine that the intention is to execute the summer clothes recommendation task, it will create a corresponding intelligent contract for the clothing recommendation service task execution. When the intelligent contract for the clothing recommendation service task is executed, the artificial intelligence service module 3 obtains data such as the user's browsing history, purchase records, and product information from the data management module 2, calls the corresponding product recommendation AI model for processing and analysis, and feeds back the recommended results output by the model to the user through speech synthesis technology.

[0075] It should be noted that during the entire operation of the above system, the system also uses ZKP technology to verify the access rights of users, homomorphic encryption and SMPC technology to ensure the privacy of model training, and records all operations based on a security audit mechanism. At the same time, for nodes that provide high-quality user data, the system gives cryptocurrency rewards through smart contracts according to the integrity of the data and its contribution to model training. For nodes that participate in model training and improve the model performance, corresponding rewards are also given. However, if it is found that a certain node attempts to tamper with data or attack the system, the smart contract will deduct its cryptocurrency and restrict its permissions, and all reward and punishment records are stored on the blockchain to ensure the fairness and transparency of the entire process.

[0076] Each module in the above decentralized personalized service system can be implemented in whole or in part by software, hardware, and their combinations. The above-mentioned modules can be embedded in the processor of the computer device in hardware form or independent of it, or stored in the memory of the computer device in software form, so that the processor can call and execute the operations corresponding to each of the above modules.

[0077] In some embodiments, as Figure 7 shown, a decentralized personalized service method is provided, and the method includes:

[0078] S11. Real-time receive user data uploaded by each user node, and encrypt and store the user data in chunks on the blockchain.

[0079] In some embodiments, the step of encrypting and storing the user data in chunks on the blockchain includes:

[0080] Based on the encryption processing of the user data of each user node, encrypted user data is obtained;

[0081] Based on the segmentation processing of each encrypted user data, a plurality of encrypted data blocks are obtained, and each encrypted data block is stored on the blockchain based on the distributed hash table technology;

[0082] Based on the content-addressable storage technology, the metadata corresponding to each encrypted data block is stored on the blockchain; the metadata includes data source, data type, timestamp, data block hash value, and access permission.

[0083] S12. Obtain the task requirements of each user node, obtain the corresponding task-related data from the blockchain based on the task requirements, and publish the task smart contract corresponding to the task requirements based on the task-related data; the task smart contract includes a model training task smart contract and a service task execution smart contract.

[0084] In some embodiments, the step of obtaining the task requirements of each user node includes:

[0085] Perform intent analysis on the task input data obtained based on the intelligent interaction interface to obtain the task requirements corresponding to the task input data.

[0086] Based on the intelligent interaction interface, feedback the task requirements to the user node that sent the task input data corresponding thereto to confirm or adjust the task requirements.

[0087] In some embodiments, the task input data includes at least one of text data, voice data, and image data; the step of performing intent analysis on the task input data obtained based on the intelligent interaction interface to obtain the task requirements corresponding to the task input data includes: performing semantic analysis on the text data based on natural language processing technology to generate text data requirements; performing recognition and conversion on the voice data based on voice recognition technology to obtain voice text data, and performing semantic analysis on the voice text data based on the natural language processing technology to generate voice data requirements; performing feature extraction and analysis on the image data based on computer vision technology to generate image data requirements; performing comprehensive requirement analysis based on the text data requirements, the voice data requirements, and / or the image data requirements to generate the task requirements.

[0088] S13. In response to the execution of the model training task smart contract, train and construct a task AI model corresponding to the task requirements based on a preset federated learning mechanism, store the task AI model in the blockchain, and feedback the execution result of the model training task smart contract to the user node.

[0089] In some embodiments, the step of training and constructing a task AI model corresponding to the task requirements based on a preset federated learning mechanism in response to the execution of the model training task smart contract includes:

[0090] Generate a model training task based on the model training requirement information in the model training task smart contract; the model training requirement information includes a model training target, a training data type, and training model parameters;

[0091] Perform task splitting and training on the model training task based on the preset federated learning mechanism to obtain multiple encrypted sub-task model parameters, and perform weighted aggregation on all the encrypted sub-task model parameters based on a first aggregation smart contract to obtain the task AI model.

[0092] In some embodiments, the step of performing task splitting and training on the model training task based on the preset federated learning mechanism to obtain multiple encrypted sub-task model parameters, and performing weighted aggregation on all the encrypted sub-task model parameters based on a first aggregation smart contract to obtain the task AI model includes:

[0093] Dispatch the model training task to the master aggregation node, and the master aggregation node splits the model training task into multiple sub-training tasks based on the preset federated learning mechanism and distributes them to different secondary aggregation nodes;

[0094] Each secondary aggregation node distributes the corresponding sub-training task to the corresponding multiple training nodes, and performs weighted aggregation on the encrypted local model parameters obtained by each training node based on the preset training smart contract to obtain the corresponding encrypted sub-task model parameters and upload them to the master aggregation node;

[0095] The master aggregation node performs weighted aggregation on the encrypted sub-task model parameters of each secondary aggregation node based on the first aggregation smart contract to obtain global model parameters;

[0096] Evaluate the global model parameters based on the preset model evaluation smart contract. When the global model parameters do not reach the model training target, update the model training task based on the global model parameters, and continue iterative training based on the updated model training task until the obtained global model parameters reach the model training target, and obtain the task AI model based on the corresponding global model parameters.

[0097] In some embodiments, the encrypted training data used by each training node to perform model training includes the encrypted user data stored in the blockchain and / or the locally stored encrypted data.

[0098] In some embodiments, the preset training smart contract includes performing model training using encrypted training data based on homomorphic encryption technology.

[0099] In some embodiments, the preset training smart contract includes adjusting the model training strategy based on the distribution difference between the local model performance index distribution of each training node and the sub-task model performance index distribution of the corresponding secondary aggregation node.

[0100] In some embodiments, the encrypted local model parameters of each training node are obtained by adding noise to the local model parameters obtained by each training node performing model training based on the preset local differential privacy mechanism.

[0101] In some embodiments, the privacy budget in the preset local differential privacy mechanism is set based on the iterative training progress, initial privacy budget, and data sensitivity of each training node; the privacy budget is inversely proportional to the iterative training progress; the privacy budget is directly proportional to the initial privacy budget and the data sensitivity.

[0102] In some embodiments, when performing weighted aggregation on the encrypted local model parameters of each training node, the aggregation weights of each training node are learned based on a preset weight allocation factor in combination with a meta-learning architecture; the preset weight allocation factor includes the data quality, data distribution, and model performance of each training node.

[0103] In some embodiments, weighted aggregation is performed on the encrypted local model parameters of each training node based on secure multi-party computation technology; and / or,

[0104] Weighted aggregation is performed on the encrypted sub-task model parameters of each sub-aggregation node based on secure multi-party computation technology.

[0105] In some embodiments, each training node uploads the encrypted local model parameters compressed based on sparse quantization technology to the corresponding sub-aggregation node; and / or,

[0106] Each sub-aggregation node uploads the encrypted sub-task model parameters compressed based on sparse quantization technology to the main aggregation node.

[0107] In some embodiments, each training node uploads the encrypted local model parameters to the corresponding sub-aggregation node based on a preset asynchronous communication mechanism, and the sub-aggregation node continuously updates the encrypted sub-task model parameters based on a preset sub-task model parameter update smart contract; and / or,

[0108] Each sub-aggregation node uploads the encrypted sub-task model parameters to the main aggregation node based on a preset asynchronous communication mechanism, and the main aggregation node continuously updates the global model parameters based on a global model parameter update smart contract.

[0109] S14. In response to the execution of the service task execution smart contract, execute the service task based on the task AI model corresponding to the task requirement, store the obtained service task execution result in the blockchain, and feedback the execution result of the service task execution smart contract to the user node.

[0110] In some embodiments, the step of executing the service task based on the task AI model corresponding to the task requirement in response to the execution of the service task execution smart contract includes:

[0111] Generate a corresponding service task based on the service requirement information in the service task execution smart contract;

[0112] Obtain corresponding service-related data based on the service task; the service-related data includes service-related user data and the corresponding task AI model;

[0113] Execute the service task based on the service-related data and a preset service execution smart contract to obtain a service task execution result.

[0114] In some embodiments, the method further includes:

[0115] Evaluating the contributions of the behaviors of each member node in real time, and generating corresponding reward and punishment measures based on the obtained behavior contribution evaluation results and a preset reward and punishment smart contract execution incentive mechanism, and storing the measures in the blockchain; the behaviors of the member nodes include data sharing, model training participation, and system maintenance.

[0116] Regularly performing security audits and vulnerability checks on all member nodes, smart contracts, data storage, and data transmission based on a preset security audit mechanism, and storing the obtained audit and inspection results in the blockchain.

[0117] For the specific limitations of the above decentralized personalized service method, reference can be made to the limitations of the decentralized personalized service system in the foregoing text, and the corresponding technical effects can also be equivalently obtained, which will not be elaborated here. It should be noted that although the steps in the above flow chart are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise clearly stated in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders.

[0118] In summary, compared with the prior art, the decentralized personalized service system and method provided by the embodiments of the present invention, through the deep integration of Web3 technology and artificial intelligence technology, combined with decentralized data storage and multi-layer encryption technology, and an improved federated learning mechanism based on privacy protection and meta-learning, give full play to the application advantages of Web3 technology and artificial intelligence technology to meet the complex requirements in the decentralized application scenario. It can not only prevent the leakage and abuse of user data during storage, training, and interaction, effectively resist various network attacks and security threats, and ensure the sustainable and stable operation of the system, but also make full use of the data resources of the distributed network to improve the accuracy and generalization ability of the constructed model, and can also accurately understand the user's behavior intention based on multi-modal input and provide reliable service feedback in a timely manner, making the interaction more natural and convenient, thereby realizing more intelligent, secure, and efficient data processing and application services, and effectively improving user satisfaction and loyalty.

[0119] Each embodiment in this specification is described in a progressive manner. For parts that are the same or similar in each embodiment, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for system embodiments, since they are basically similar to method embodiments, the description is relatively simple. For related parts, reference can be made to the description of the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as within the scope described in this specification.

[0120] The above-described embodiments merely represent several preferred embodiments of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be pointed out that for those of ordinary skill in the art in this technical field, without departing from the technical principle of the present invention, several improvements and substitutions can be made, and these improvements and substitutions should also be regarded as within the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the protection scope of the claims.

Claims

1. A decentralized personalized service system, characterized in that: The personalized service system includes an intelligent interaction module, a data management module and an artificial intelligence service module; The intelligent interaction module is used to receive user data uploaded by each user node, and transmit the user data to the data management module based on the data storage request; it is also used to obtain the task requirements of each user node, obtain the task-related data corresponding to the task requirements from the data management module based on the data call request, and publish the task smart contract corresponding to the task requirements based on the task-related data, and in response to the completion of the execution of the task smart contract, feedback the execution result of the task smart contract to the user node; the task smart contract includes a model training task smart contract and a service task execution smart contract; The data management module is used to store the user data transmitted by the intelligent interaction module in encrypted blocks in the blockchain based on the data storage request; and is also used to transmit the task-related data corresponding to the task requirement to the intelligent interaction module based on the data call request; The artificial intelligence service module is used to respond to the execution of the model training task smart contract, train and construct a task AI model corresponding to the task requirements based on a preset federated learning mechanism, and store the task AI model in the blockchain; it is also used to respond to the execution of the service task execution smart contract, execute the service task based on the task AI model corresponding to the task requirements, and store the obtained service task execution results in the blockchain.

2. The decentralized personalized service system according to claim 1, characterized in that: The personalized service system also includes an incentive management module and a security protection module; The incentive management module is used to evaluate the contribution of each member node behavior in real time, and based on the obtained behavior contribution evaluation results and the preset reward and punishment smart contract execution incentive mechanism, generate corresponding reward and punishment measures and store them in the blockchain; The member node behaviors include data sharing, model training participation and system maintenance; The security protection module is used to regularly perform security audits and vulnerability checks on member nodes, smart contracts, data storage and data transmission in the personalized service system based on a preset security audit mechanism, and store the obtained audit results in the blockchain.

3. The decentralized personalized service system according to claim 1, characterized in that: The intelligent interaction module includes a demand analysis unit and a demand confirmation unit; The demand analysis unit is used to perform intention analysis on the task input data obtained based on the intelligent interactive interface to obtain the task demand corresponding to the task input data; The demand confirmation unit is used to feed back the task demand to the user node corresponding to the task input data based on the intelligent interactive interface, so as to confirm or adjust the task demand.

4. The decentralized personalized service system according to claim 3, characterized in that: The task input data includes at least one of text data, voice data and image data; the demand analysis unit includes a text data analysis unit, a voice data analysis unit, an image data analysis unit and a comprehensive analysis unit; The text data analysis unit is used to generate text data requirements by performing semantic analysis on the text data based on natural language processing technology; The voice data analysis unit is used to perform recognition conversion on the voice data based on the voice recognition technology to obtain voice text data, and to perform semantic analysis on the voice text data based on the natural language processing technology to generate voice data requirements; The image data analysis unit is used to perform feature extraction and analysis on the image data based on computer vision technology to generate image data requirements; The comprehensive analysis unit is used to generate the task requirement based on a comprehensive requirement analysis of the text data requirement, the voice data requirement and / or the image data requirement.

5. The decentralized personalized service system according to claim 1, characterized in that: The data management module includes an encryption block unit and a data storage unit; The encryption block division unit is used to obtain encrypted user data based on the encryption processing performed on the user data of each user node, and obtain multiple encrypted data blocks based on the segmentation processing performed on each encrypted user data; The data storage unit is used to store each encrypted data block in the blockchain based on the distributed hash table technology, and store the metadata corresponding to each encrypted data block in the blockchain based on the content-addressed storage technology; the metadata includes data source, data type, timestamp, data block hash value and access rights.

6. The decentralized personalized service system according to claim 1, characterized in that: The artificial intelligence service module includes a model training unit and a model execution unit; The model training unit is used to generate a model training task based on the model training requirement information in the model training task smart contract, perform task splitting training on the model training task based on the preset federated learning mechanism, obtain multiple encrypted subtask model parameters, and perform weighted aggregation on all encrypted subtask model parameters based on the first aggregation smart contract to obtain the task AI model; the model training requirement information includes model training objectives, training data types and training model parameters; The model execution unit is used to generate a corresponding service task based on the service demand information in the service task execution smart contract, obtain service-related data corresponding to the service task based on the interaction with the data management module, and execute the service task based on the service-related data and the preset service execution smart contract to obtain the service task execution result; the service-related data includes service-related user data and the corresponding task AI model.

7. The decentralized personalized service system according to claim 6, characterized in that: The model training unit includes a task splitting subunit, a subtask training subunit, a model aggregation subunit and a model evaluation subunit; The task splitting subunit is used to send the model training task to the main aggregation node, and the main aggregation node splits the model training task into multiple sub-training tasks based on the preset federated learning mechanism and distributes them to different secondary aggregation nodes; The subtask training subunit is used to distribute the corresponding sub-training tasks to the corresponding multiple training nodes by each secondary aggregation node, and perform weighted aggregation on the encrypted local model parameters obtained by each training node based on the preset training smart contract execution model training, and obtain the corresponding encrypted subtask model parameters and upload them to the main aggregation node; The model aggregation subunit is used for the main aggregation node to perform weighted aggregation on the encrypted subtask model parameters of each secondary aggregation node based on the first aggregation smart contract to obtain a global model parameter; The model evaluation subunit is used to evaluate the global model parameters based on a preset model evaluation smart contract. When the global model parameters do not reach the model training target, the model training task is updated based on the global model parameters, and iterative training is continued based on the updated model training task until the global model parameters obtained reach the model training target, and the task AI model is obtained based on the corresponding global model parameters.

8. The decentralized personalized service system according to claim 7, characterized in that: The encrypted training data used by each training node to perform model training includes encrypted data obtained based on the interaction with the data management module and / or locally stored encrypted data.

9. The decentralized personalized service system according to claim 7, characterized in that: The preset training smart contract includes executing model training based on homomorphic encryption technology and using encrypted training data.

10. The decentralized personalized service system according to claim 7, characterized in that: The preset training smart contract includes adjusting the model training strategy based on the distribution difference between the local model performance indicator distribution of each training node and the subtask model performance indicator distribution of the corresponding secondary aggregation node.

11. The decentralized personalized service system according to claim 7, characterized in that: The encrypted local model parameters of each training node are obtained by adding noise to the local model parameters obtained by performing model training on each training node based on a preset local differential privacy mechanism.

12. The decentralized personalized service system according to claim 11, characterized in that: The privacy budget in the preset local differential privacy mechanism is set based on the iterative training progress, initial privacy budget and data sensitivity of each training node; the privacy budget is inversely proportional to the iterative training progress; the privacy budget is proportional to the initial privacy budget and the data sensitivity.

13. The decentralized personalized service system according to claim 7, characterized in that: Performing weighted aggregation of encrypted local model parameters of each training node based on secure multi-party computing technology; and / or, The encryption subtask model parameters of each sub-aggregation node are weightedly aggregated based on secure multi-party computing technology.

14. The decentralized personalized service system according to claim 7 or 13, characterized in that: The aggregation weight of each training node in the weighted aggregation of the encrypted local model parameters of each training node is obtained based on preset weight allocation factors combined with meta-learning architecture learning; the preset weight allocation factors include data quality, data distribution and model performance of each training node.

15. The decentralized personalized service system according to claim 7, characterized in that: Each training node uploads the encrypted local model parameters compressed based on sparse quantization technology to the corresponding secondary aggregation node; and / or, Each secondary aggregation node uploads the encrypted subtask model parameters compressed based on the sparse quantization technology to the main aggregation node.

16. The decentralized personalized service system according to claim 7 or 15, characterized in that: Each training node uploads the encrypted local model parameters to the corresponding sub-aggregation node based on a preset asynchronous communication mechanism, and the sub-aggregation node continuously updates the encrypted sub-task model parameters based on a preset sub-task model parameter update smart contract; and / or, Each secondary aggregation node will upload the encrypted subtask model parameters to the main aggregation node based on a preset asynchronous communication mechanism, and the main aggregation node will continuously update the global model parameters based on the global model parameter update smart contract.

17. A decentralized personalized service method, characterized in that: The method comprises: Receive user data uploaded by each user node in real time, and encrypt and block the user data and store it in the blockchain; Obtaining task requirements of each user node, obtaining corresponding task-related data from the blockchain based on the task requirements, and publishing a task smart contract corresponding to the task requirements based on the task-related data; the task smart contract includes a model training task smart contract and a service task execution smart contract; In response to the execution of the model training task smart contract, a task AI model corresponding to the task requirement is trained and constructed based on a preset federated learning mechanism, and the task AI model is stored in the blockchain, and the execution result of the model training task smart contract is fed back to the user node; In response to the execution of the service task execution smart contract, the service task is executed based on the task AI model corresponding to the task requirement, and the obtained service task execution result is stored in the blockchain, and the execution result of the service task execution smart contract is fed back to the user node.

18. The decentralized personalized service method according to claim 17, characterized in that: The method further comprises: Conduct real-time contribution evaluation of each member node behavior, and generate corresponding reward and punishment measures based on the obtained behavior contribution evaluation results and the preset reward and punishment smart contract execution incentive mechanism, and store them in the blockchain; the member node behavior includes data sharing, model training participation and system maintenance; Based on the preset security audit mechanism, security audits and vulnerability checks are regularly performed on all member nodes, smart contracts, data storage, and data transmission, and the audit and check results obtained are stored in the blockchain.

19. The decentralized personalized service method according to claim 17, characterized in that: The step of obtaining the task requirements of each user node includes: Performing intention analysis on task input data acquired based on the intelligent interactive interface to obtain task requirements corresponding to the task input data; Based on the intelligent interactive interface, the task requirement is fed back to the user node that sends the task input data to confirm or adjust the task requirement.

20. The decentralized personalized service method according to claim 19, characterized in that: The task input data includes at least one of text data, voice data and image data; the step of performing intention analysis on the task input data obtained based on the intelligent interactive interface to obtain the task requirements corresponding to the task input data includes: Generate text data requirements by performing semantic analysis on the text data based on natural language processing technology; Based on the speech recognition technology, the speech data is recognized and converted to obtain speech text data, and based on the natural language processing technology, the speech text data is semantically analyzed to generate speech data requirements; Perform feature extraction and analysis on the image data based on computer vision technology to generate image data requirements; The task requirement is generated based on a comprehensive requirement analysis of the text data requirement, the voice data requirement and / or the image data requirement.

21. The decentralized personalized service method according to claim 17, characterized in that: The step of encrypting and storing the user data in blocks on the blockchain comprises: Based on the encryption processing performed on the user data of each user node, encrypted user data is obtained; Based on the segmentation process of each encrypted user data, a plurality of encrypted data blocks are obtained, and each encrypted data block is stored in the blockchain based on the distributed hash table technology; The metadata corresponding to each encrypted data block is stored in the blockchain based on content-addressable storage technology; the metadata includes data source, data type, timestamp, data block hash value and access rights.

22. The decentralized personalized service method according to claim 17, characterized in that: In response to the execution of the model training task smart contract, the step of training and constructing a task AI model corresponding to the task requirement based on a preset federated learning mechanism includes: Generate a model training task based on the model training requirement information in the model training task smart contract; the model training requirement information includes model training objectives, training data types and training model parameters; Based on the preset federated learning mechanism, the model training task is split and trained to obtain multiple encrypted sub-task model parameters, and all encrypted sub-task model parameters are weighted aggregated based on the first aggregation smart contract to obtain the task AI model.

23. The decentralized personalized service method according to claim 22, characterized in that: The step of performing task splitting training on the model training task based on the preset federated learning mechanism to obtain multiple encrypted subtask model parameters, and performing weighted aggregation on all encrypted subtask model parameters based on the first aggregation smart contract to obtain the task AI model includes: Sending the model training task to the main aggregation node, which splits the model training task into multiple sub-training tasks based on the preset federated learning mechanism and distributes them to different sub-aggregation nodes; Each secondary aggregation node distributes the corresponding sub-training tasks to the corresponding multiple training nodes, and performs weighted aggregation on the encrypted local model parameters obtained by each training node based on the preset training smart contract to execute model training, and obtains the corresponding encrypted sub-task model parameters and uploads them to the main aggregation node; The main aggregation node performs weighted aggregation on the encrypted subtask model parameters of each secondary aggregation node based on the first aggregation smart contract to obtain a global model parameter; The global model parameters are evaluated based on the preset model evaluation smart contract. When the global model parameters do not reach the model training target, the model training task is updated based on the global model parameters, and iterative training is continued based on the updated model training task until the global model parameters reach the model training target, and the task AI model is obtained based on the corresponding global model parameters.

24. The decentralized personalized service method according to claim 23, characterized in that: The encrypted training data used by each training node to perform model training includes the encrypted user data stored in the blockchain and / or the locally stored encrypted data.

25. The decentralized personalized service method according to claim 23, characterized in that: The preset training smart contract includes executing model training based on homomorphic encryption technology and using encrypted training data.

26. The decentralized personalized service method according to claim 23 or 25, characterized in that: The preset training smart contract includes adjusting the model training strategy based on the distribution difference between the local model performance indicator distribution of each training node and the subtask model performance indicator distribution of the corresponding secondary aggregation node.

27. The decentralized personalized service method according to claim 23, characterized in that: The encrypted local model parameters of each training node are obtained by adding noise to the local model parameters obtained by performing model training on each training node based on a preset local differential privacy mechanism.

28. The decentralized personalized service method according to claim 27, characterized in that: The privacy budget in the preset local differential privacy mechanism is set based on the iterative training progress, initial privacy budget and data sensitivity of each training node; the privacy budget is inversely proportional to the iterative training progress; the privacy budget is proportional to the initial privacy budget and the data sensitivity.

29. The decentralized personalized service method according to claim 23, characterized in that: Performing weighted aggregation of encrypted local model parameters of each training node based on secure multi-party computing technology; and / or, The encryption subtask model parameters of each sub-aggregation node are weightedly aggregated based on secure multi-party computing technology.

30. The decentralized personalized service method according to claim 23 or 29, characterized in that: The aggregation weight of each training node in the weighted aggregation of the encrypted local model parameters of each training node is obtained based on preset weight allocation factors combined with meta-learning architecture learning; the preset weight allocation factors include data quality, data distribution and model performance of each training node.

31. The decentralized personalized service method according to claim 23, characterized in that: Each training node uploads the encrypted local model parameters compressed based on sparse quantization technology to the corresponding secondary aggregation node; and / or, Each secondary aggregation node uploads the encrypted subtask model parameters compressed based on the sparse quantization technology to the main aggregation node.

32. The decentralized personalized service method according to claim 23 or 31, characterized in that: Each training node uploads the encrypted local model parameters to the corresponding sub-aggregation node based on a preset asynchronous communication mechanism, and the sub-aggregation node continuously updates the encrypted sub-task model parameters based on a preset sub-task model parameter update smart contract; and / or, Each secondary aggregation node will upload the encrypted subtask model parameters to the main aggregation node based on a preset asynchronous communication mechanism, and the main aggregation node will continuously update the global model parameters based on the global model parameter update smart contract.

33. The decentralized personalized service method according to claim 17, characterized in that: The step of executing the service task based on the task AI model corresponding to the task requirement in response to the execution of the service task execution smart contract includes: Generate a corresponding service task based on the service demand information in the service task execution smart contract; Acquire corresponding service-related data based on the service task; the service-related data includes service-related user data and a corresponding task AI model; Based on the service-related data and the preset service execution smart contract, the service task is executed to obtain the service task execution result.

Images