Big data privacy protection modeling method and system based on federated learning and block chain

Abstract

The invention discloses a big data privacy protection modeling method and system based on federated learning and a block chain, and relates to the technical field of privacy protection and joint modeling. According to the method, homomorphic encryption, differential privacy, federated learning, secure multi-party computing and block chain technologies are fused, big data privacy protection and joint modeling are realized, encryption and dimensionality reduction are performed on original data through homomorphic encryption and differential privacy, an encrypted training sample of secure privacy is generated, a local model is trained on an encrypted data set through federated learning, and a big data privacy protection result is obtained. The method comprises the following steps: calculating aggregation parameters by using security multiple parties, constructing a verification network in combination with a block chain, ensuring credibility and integrity of model training, and finally, adding noise optimization performance for a global model by using differential privacy, testing generalization ability through cross validation, and determining a deployable privacy protection joint learning model, thereby breaking traditional data islands, promoting cross-mechanism data cooperation, and improving the privacy protection performance. Big data values are released, and data protection regulations and privacy requirements are met.

Description

Technical Field

[0001] This invention relates to the field of privacy protection and collaborative modeling technology, specifically to a big data privacy protection modeling method and system based on federated learning and blockchain. Background Technology

[0002] With the deepening of digital transformation, big data has become a core driving force for innovation and development across industries, and the demand for cross-institutional and cross-domain data collaboration is becoming increasingly urgent. However, the phenomenon of data silos severely restricts the full release of data value. All parties involved hope to obtain better model results through data sharing, but they also worry that sensitive data will be leaked or misused during the sharing process.

[0003] Traditional centralized machine learning methods require raw data to be aggregated into a unified platform for processing. This model is inadequate when faced with strict data protection regulations and privacy requirements. Some existing privacy protection solutions often adopt a single-point architecture, which has the risks of centralization and single point of failure. At the same time, they lack effective data integrity verification mechanisms and it is difficult to establish trust relationships among the participants. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for big data privacy protection modeling based on federated learning and blockchain, which realizes privacy protection and joint modeling of big data.

[0005] The objective of this invention can be achieved through the following technical solutions: This application provides a privacy-preserving modeling method for big data based on federated learning and blockchain, including the following steps: The original training data of each participating node is encrypted using a homomorphic encryption algorithm. An encrypted dataset is generated based on a preset encryption key and then dimensionality reduction is performed to obtain encrypted training samples that meet privacy protection requirements. The local model parameters are trained on each encrypted dataset using a federated learning framework. The local model weight vector is calculated using the gradient descent algorithm to obtain the encrypted model parameters to be aggregated. A secure multi-party computation protocol is used to aggregate the encrypted model parameters of each node. The parameters are then transmitted to other collaborating nodes in fragments through a secret sharing mechanism. The global model parameters are calculated based on the predefined aggregation weights to obtain a joint model that integrates the features of multi-party data. Obtain the digital signature information of each participating node and build a blockchain verification network. Generate unique identifiers for model parameters and training process based on hash algorithm to determine the credibility and integrity of the training process. The parameter verification rules are automatically executed by smart contracts to identify malicious node behavior based on preset anomaly detection thresholds, thereby obtaining a reliable set of model parameters that have been verified by trust. By adding calibration noise to the aggregated global model using differential privacy technology, and controlling the noise intensity according to the privacy budget allocation strategy, a final collaborative model that meets privacy protection standards is obtained. The collaborative training results are quality-checked based on model performance evaluation metrics. The generalization ability of the model on data at each node is tested using cross-validation methods to determine deployable privacy-preserving joint learning models.

[0006] Furthermore, encrypted training samples that meet privacy protection requirements are obtained, specifically including: The original training data of each participating node is encrypted using a homomorphic encryption algorithm. An encrypted dataset is generated based on a preset encryption key. Principal component analysis is then performed on the encrypted dataset to reduce its dimensionality, thereby obtaining the encrypted feature set after dimensionality reduction and determining the low-dimensional encrypted feature set. When the dimension of the low-dimensional encrypted feature set meets the preset threshold, a clustering algorithm is used to group the low-dimensional encrypted feature set to obtain encrypted sample groups. When the dimension does not meet the preset threshold, the principal component analysis parameters are adjusted and the low-dimensional encrypted feature set is obtained again. Based on the encrypted sample grouping, noise is added using a differential privacy algorithm to obtain the encrypted training samples with added noise, and the encrypted training samples that meet the privacy protection requirements are determined. Feature distribution information is extracted from the encrypted training samples that meet the privacy protection requirements, and it is judged whether the feature distribution information conforms to the preset statistical characteristics to obtain the feature distribution verification result. If the feature distribution verification result meets the preset statistical characteristics, then the metadata description of the encrypted training sample is generated based on the feature distribution information, and the metadata of the encrypted training sample is obtained. If it does not meet the requirements, then the noise parameters of the differential privacy algorithm are adjusted, and the encrypted training sample is obtained again.

[0007] Based on the encrypted training sample metadata, a data verification tool is used to verify the integrity and consistency of the metadata, thus obtaining the final encrypted training sample that meets privacy protection requirements.

[0008] Furthermore, data validation tools are used to verify the integrity and consistency of metadata, specifically including: A consistency check algorithm is used to verify the metadata description of the encrypted training samples and obtain the consistency check result. Then, an integrity check algorithm is used to check the integrity of the metadata description and obtain the integrity check result. Based on the integrity verification results, a feature extraction template is generated using the metadata description of the encrypted training samples. The feature extraction template is then obtained, and features are extracted from the encrypted training samples to obtain the extracted feature set.

[0009] If the distribution characteristics of the feature set meet the preset statistical criteria, a classification algorithm is used to group the feature set and obtain the encrypted sample set after grouping. If it does not meet the criteria, the feature extraction template parameters are adjusted and the feature set is obtained again.

[0010] Based on the grouped encrypted sample set, generate metadata tags for the sample set, obtain the encrypted sample tag set, and verify the encrypted sample tag set through a consistency check algorithm.

[0011] Furthermore, the parameters of the encryption model to be aggregated are obtained, specifically including: The federated learning framework obtains training data from various encrypted datasets, updates the weight vector of the local model using the gradient descent algorithm, and obtains the local model weights. When the training loss function value of the local model weights is lower than the preset convergence threshold, the local training is considered complete, and the encrypted model parameters are obtained. Based on the encrypted model parameters, a secure multi-party computation protocol is used to aggregate the parameters of each local model to obtain the global model parameters, update the weight vector of each local model, and obtain the updated local model. When the loss function value of the updated local model is lower than the global convergence threshold, the global model is determined to have converged, and the final global model is obtained. Then, differential privacy technology is used to perturb the model parameters to obtain privacy-preserving global model parameters, which are then distributed to each local node to update the local model and obtain the final encrypted model parameters to be aggregated.

[0012] Furthermore, a joint model integrating features from multiple sources is obtained, specifically including: A secure multi-party computation protocol is adopted to obtain encrypted model parameters from each node, perform parameter sharding to obtain a sharded parameter set, and transmit the sharded parameter set to the cooperating node network through a secret sharing mechanism to ensure that each node receives the complete sharded parameters. Once the receiving node verifies the integrity of the sharding parameters, it performs a weighted aggregation operation based on the predefined aggregation weights to obtain preliminary global model parameters. Then, it adds noise using a differential privacy algorithm to obtain privacy-enhanced global model parameters. Through a distributed computing architecture, privacy-enhancing global model parameters are distributed to each collaborating node, and the updated local model parameters of each node are obtained. When the deviation between the local model parameters and the global model parameters is less than a preset threshold, parameter fusion is performed to obtain a joint model that integrates the features of multiple data sources.

[0013] Furthermore, the reliability and integrity of the training process are determined, specifically including: Obtain the digital signatures of each participating node, verify the signatures using an asymmetric encryption algorithm, and obtain the signature verification result. If the signature verification result passes, extract the model parameters and training process data, generate a unique identifier using the SHA-256 hash algorithm, and obtain the unique identifier value. Based on the unique identifier value, the transaction records in the blockchain verification network are constructed, and the validity of the transaction is confirmed by the consensus algorithm to obtain the transaction confirmation status. When the transaction confirmation status is valid, the model parameter update record is written into the distributed ledger to obtain the ledger update status. Based on the ledger update status, the consistency between the parameter update records in the distributed ledger and the original training process data is verified. When the consistency verification result passes, a trustworthiness confirmation flag and an integrity confirmation flag are generated to obtain the trustworthiness and integrity status of the training process. These are used to update the node trust value in the blockchain network to obtain the node trust value update result.

[0014] Furthermore, a reliable set of model parameters that has been verified through trust is obtained, specifically including: The system obtains parameter data submitted by nodes through smart contracts, verifies the parameters using preset verification rules, and obtains a preliminary set of verified parameters. If there are parameters in the preliminary set of verified parameters that deviate from the normal range, they are marked as abnormal parameters, thus identifying potential malicious nodes. Based on the marked abnormal parameters, the penalty mechanism in the smart contract is triggered to restrict the parameter submission permissions of malicious nodes, resulting in an updated parameter set. The updated parameter set is then verified a second time using a trust verification algorithm to obtain a trust score and determine the parameter set with high trust. The federated learning algorithm aggregates the set of high-trust parameters to generate global model parameters. When the performance index of the global model parameters does not reach the preset threshold, the scoring criteria of the trust verification algorithm is adjusted, and a second verification is performed to obtain an optimized set of trust parameters. The verification rules in the smart contract are then updated to generate the final set of reliable model parameters.

[0015] Furthermore, a final collaborative model that meets privacy protection standards is obtained, specifically including: The data distribution characteristics of the aggregated global model are obtained by differential privacy technology. The Laplace mechanism is used to add calibration noise to the model parameters to obtain an initial noise model. Then, according to the privacy budget allocation strategy, the noise intensity parameter is extracted from the initial noise model, and the noise intensity is adjusted by the exponential decay method to obtain an optimized noise model. When the privacy loss of the optimized noise model exceeds the preset threshold, the noise parameters are reconfigured through gradient pruning to obtain the adjusted noise model. Then, its differential privacy protection level is obtained, and the privacy loss is calculated through the ε-differential privacy formula to determine whether the preset threshold is met, thus obtaining a privacy-compliant model. Model parameters are extracted from the privacy compliance model, and a federated learning framework is used for multi-party collaborative updates to obtain a collaborative update model. Then, the performance indicators of the global model are obtained through the aggregated parameters of the collaborative update model to determine whether the privacy protection standards are met, thus obtaining the final collaborative model.

[0016] Furthermore, deployable privacy-preserving joint learning models are identified, specifically including: The training and validation subsets are obtained from the data of each node through cross-validation to obtain the partitioned dataset. The collaborative model is trained using the K-fold cross-validation method to obtain the prediction results of each fold. The performance metrics, including accuracy and loss value, are obtained from the prediction results to determine the model's generalization ability on data at each node. When the accuracy reaches a preset threshold, the model parameters of each node are aggregated through a joint learning framework to obtain the global model. A differential privacy algorithm is used to add noise to obtain a privacy-preserving model. The performance metrics of the privacy-preserving model are tested using a validation set to determine whether it meets the deployability requirements. If the performance metrics meet a preset threshold, a deployable privacy-preserving joint learning model is determined.

[0017] This application provides a big data privacy-preserving modeling system based on federated learning and blockchain, used to implement a big data privacy-preserving modeling method based on federated learning and blockchain, including: The data encryption and dimensionality reduction module performs homomorphic encryption and dimensionality reduction on the original data of the participating nodes to generate encrypted training samples. The federated learning training module trains a local model on an encrypted dataset using a federated learning framework, and uses the gradient descent algorithm to calculate the weights of the local model, thus enabling the training and updating of the local model. The multi-party computation and model aggregation module uses a secure multi-party computation protocol to aggregate encrypted model parameters from each node, and calculates global model parameters through a secret sharing mechanism and predefined weights to generate a joint model. The verification and trust confirmation module constructs a blockchain verification network to perform hash identification and signature verification on model parameters and training process, and writes parameter update records into a distributed ledger. The smart contract and trust verification module automatically executes parameter verification rules through smart contracts, identifies malicious node behavior and triggers a penalty mechanism, performs secondary verification of the parameter set, and generates a reliable model parameter set. The differential privacy model generation module adds calibration noise to the aggregated global model, controls the noise intensity according to the privacy budget, adjusts the noise parameters, and generates a privacy-preserving collaborative model. The model evaluation and deployment module performs quality checks on the collaborative training results based on performance evaluation metrics, tests the model's generalization ability through cross-validation, and determines deployable privacy-preserving joint learning models.

[0018] The beneficial effects of this invention are as follows: Homomorphic encryption, dimensionality reduction, and differential privacy algorithms are used to generate encrypted training samples and verify the integrity and consistency of metadata. This effectively solves the problem of sensitive data leakage or misuse during data sharing, ensures the secure sharing and use of data under privacy protection, provides a secure and reliable data foundation for subsequent joint modeling, and promotes the full release of the value of big data. By training a local model using a federated learning framework, aggregating encrypted model parameters using a secure multi-party computation protocol, constructing a verification network using blockchain technology and recording parameter updates, and identifying malicious node behavior through smart contracts, this approach addresses the shortcomings of traditional centralized machine learning methods under data protection regulations and privacy requirements. It establishes trust relationships among participants, ensures the credibility and integrity of the model training process, and improves the reliability and security of the model. By adding calibration noise to the global model using differential privacy technology and controlling the noise intensity, a final collaborative model that meets privacy protection standards is generated. The model's generalization ability is tested through cross-validation, and a deployable privacy-preserving joint learning model is determined. This solves the problems of existing privacy protection schemes lacking data integrity verification mechanisms and difficulty in establishing trust relationships among participants. It achieves the optimization of model performance and ensures its generalization ability while protecting data privacy, ultimately resulting in a deployable privacy-preserving joint learning model. This promotes the realization of cross-institutional and cross-domain data collaboration and facilitates the full release of the value of big data. Attached Figure Description

[0019] To better understand and implement this application, the technical solution is described in detail below with reference to the accompanying drawings.

[0020] Figure 1 A flowchart illustrating the big data privacy protection modeling method based on federated learning and blockchain provided in Embodiment 1 of this application; Figure 2 A flowchart illustrating the process for determining the credibility and integrity of the training process in the big data privacy protection modeling method based on federated learning and blockchain provided in Embodiment 1 of this application; Figure 3 A schematic diagram illustrating the process of obtaining a reliable set of model parameters verified by trust using the big data privacy protection modeling method based on federated learning and blockchain provided in Embodiment 1 of this application; Figure 4 This is a schematic diagram of the structure of the big data privacy protection modeling system based on federated learning and blockchain provided in Embodiment 2 of this application. Detailed Implementation

[0021] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, exemplary embodiments will be described in detail below, examples of which are illustrated in the accompanying drawings. In the following description relating to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of methods and systems consistent with some aspects of this application as detailed in the appended claims.

[0022] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0023] The following detailed description of the specific implementation methods, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided in detail.

[0024] Example 1

[0025] Please see Figures 1-3 This embodiment provides a big data privacy protection modeling method based on federated learning and blockchain, including the following steps: S1. The original training data of each participating node is encrypted using a homomorphic encryption algorithm. An encrypted dataset is generated based on the preset encryption key, and then dimensionality reduction is performed to obtain encrypted training samples that meet the privacy protection requirements. Furthermore, encrypted training samples that meet privacy protection requirements are obtained, specifically including: The original training data of each participating node is encrypted using a homomorphic encryption algorithm. An encrypted dataset is generated based on a preset encryption key. Principal component analysis is then performed on the encrypted dataset to reduce its dimensionality, thereby obtaining the encrypted feature set after dimensionality reduction and determining the low-dimensional encrypted feature set. When the dimension of the low-dimensional encrypted feature set meets the preset threshold, a clustering algorithm is used to group the low-dimensional encrypted feature set to obtain encrypted sample groups. When the dimension does not meet the preset threshold, the principal component analysis parameters are adjusted and the low-dimensional encrypted feature set is obtained again. Based on the encrypted sample grouping, noise is added using a differential privacy algorithm to obtain the encrypted training samples with added noise, and the encrypted training samples that meet the privacy protection requirements are determined. Feature distribution information is extracted from the encrypted training samples that meet the privacy protection requirements, and it is judged whether the feature distribution information conforms to the preset statistical characteristics to obtain the feature distribution verification result. If the feature distribution verification result meets the preset statistical characteristics, then the metadata description of the encrypted training sample is generated based on the feature distribution information, and the metadata of the encrypted training sample is obtained. If it does not meet the requirements, then the noise parameters of the differential privacy algorithm are adjusted, and the encrypted training sample is obtained again.

[0026] Based on the encrypted training sample metadata, a data verification tool is used to verify the integrity and consistency of the metadata, thus obtaining the final encrypted training sample that meets privacy protection requirements.

[0027] Specifically, the original data is encrypted using homomorphic encryption and differential privacy algorithms, and then combined with principal component analysis for dimensionality reduction and feature distribution verification to generate encrypted training samples that meet privacy protection requirements. This process not only ensures the security of data during transmission and use, preventing the leakage of sensitive information, but also verifies the integrity and consistency of metadata through data verification tools, further guaranteeing the quality and reliability of the data. The resulting encrypted training samples meet both privacy protection requirements and possess good data characteristics, providing a high-quality data foundation for subsequent joint modeling and effectively solving the dual problems of privacy protection and data quality in data sharing.

[0028] Furthermore, data validation tools are used to verify the integrity and consistency of metadata, specifically including: A consistency check algorithm is used to verify the metadata description of the encrypted training samples and obtain the consistency check result. Then, an integrity check algorithm is used to check the integrity of the metadata description and obtain the integrity check result. Based on the integrity verification results, a feature extraction template is generated using the metadata description of the encrypted training samples. The feature extraction template is then obtained, and features are extracted from the encrypted training samples to obtain the extracted feature set.

[0029] If the distribution characteristics of the feature set meet the preset statistical criteria, a classification algorithm is used to group the feature set and obtain the encrypted sample set after grouping. If it does not meet the criteria, the feature extraction template parameters are adjusted and the feature set is obtained again.

[0030] Based on the grouped encrypted sample set, generate metadata tags for the sample set, obtain the encrypted sample tag set, and verify the encrypted sample tag set through a consistency check algorithm.

[0031] Specifically, by verifying the encrypted sample tag set through a consistency check algorithm, the integrity and consistency of the data are further ensured, effectively improving the reliability and trustworthiness of the entire data processing flow.

[0032] S2. Train local model parameters on each encrypted dataset using a federated learning framework, calculate local model weight vectors using gradient descent algorithm, and determine local training completion when the training loss function value is lower than the convergence threshold, thus obtaining the encrypted model parameters to be aggregated. Furthermore, the parameters of the encryption model to be aggregated are obtained, specifically including: The federated learning framework obtains training data from various encrypted datasets, updates the weight vector of the local model using the gradient descent algorithm, and obtains the local model weights. When the training loss function value of the local model weights is lower than the preset convergence threshold, the local training is considered complete, and the encrypted model parameters are obtained. Based on the encrypted model parameters, a secure multi-party computation protocol is used to aggregate the parameters of each local model to obtain the global model parameters, update the weight vector of each local model, and obtain the updated local model. When the loss function value of the updated local model is lower than the global convergence threshold, the global model is determined to have converged, and the final global model is obtained. Then, differential privacy technology is used to perturb the model parameters to obtain privacy-preserving global model parameters, which are then distributed to each local node to update the local model and obtain the final encrypted model parameters to be aggregated.

[0033] Specifically, a federated learning framework is used to train local models on encrypted datasets, and gradient descent is used to update weights to ensure the local model training is complete. A secure multi-party computation protocol is then used to aggregate the local model parameters to generate global model parameters. Differential privacy technology is used to perturb the global model parameters to protect privacy. This process addresses the privacy leakage risks caused by the need for data to be aggregated on a unified platform in traditional centralized machine learning, as well as the centralization risks and single-point failure issues inherent in single-point architectures. Furthermore, federated learning enables distributed training, improving the model's generalization ability and privacy protection level, ensuring that all participating parties can collaboratively train a high-performance global model while protecting data privacy.

[0034] S3. A secure multi-party computation protocol is used to aggregate the encrypted model parameters of each node. The parameters are then transmitted to other collaborating nodes in fragments through a secret sharing mechanism. The global model parameters are calculated based on the predefined aggregation weights to obtain a joint model that integrates the features of multi-party data. Furthermore, a joint model integrating features from multiple sources is obtained, specifically including: A secure multi-party computation protocol is adopted to obtain encrypted model parameters from each node, perform parameter sharding to obtain a sharded parameter set, and transmit the sharded parameter set to the cooperating node network through a secret sharing mechanism to ensure that each node receives the complete sharded parameters. Once the receiving node verifies the integrity of the sharding parameters, it performs a weighted aggregation operation based on the predefined aggregation weights to obtain preliminary global model parameters. Then, it adds noise using a differential privacy algorithm to obtain privacy-enhanced global model parameters. Through a distributed computing architecture, privacy-enhancing global model parameters are distributed to each collaborating node, and the updated local model parameters of each node are obtained. When the deviation between the local model parameters and the global model parameters is less than a preset threshold, parameter fusion is performed to obtain a joint model that integrates the features of multiple data sources.

[0035] Specifically, this paper addresses the challenge of effectively aggregating multi-party data features to improve model performance while protecting the data privacy of all participants in distributed collaborative modeling. Furthermore, it enhances privacy protection by adding noise to global model parameters using a differential privacy algorithm, ensuring data security and privacy during the collaborative process. Additionally, it distributes privacy-enhanced global model parameters through a distributed computing architecture and performs parameter fusion when the deviation between local and global model parameters is less than a preset threshold, achieving efficient model updates and optimization, and improving the model's generalization ability and overall performance.

[0036] S4. Obtain the digital signature information of each participating node and build a blockchain verification network. Generate unique identifiers for model parameters and training process according to the hash algorithm. When the signature verification is successful, write the parameter update record into the distributed ledger to determine the credibility and integrity of the training process. Furthermore, the reliability and integrity of the training process are determined, specifically including: S41. Obtain the digital signature of each participating node, verify the signature using an asymmetric encryption algorithm, and obtain the signature verification result. If the signature verification result passes, extract the model parameters and training process data, generate a unique identifier using the SHA-256 hash algorithm, and obtain the unique identifier value. S42. Based on the unique identifier value, construct the transaction record in the blockchain verification network, use the consensus algorithm to confirm the validity of the transaction, and obtain the transaction confirmation status. When the transaction confirmation status is valid, write the model parameter update record into the distributed ledger to obtain the ledger update status. S43. Based on the ledger update status, verify the consistency between the parameter update records in the distributed ledger and the original training process data to obtain the consistency verification result. When the consistency verification result passes, generate the trustworthiness confirmation flag and the integrity confirmation flag to obtain the trustworthiness and integrity status of the training process, which is used to update the node trust value in the blockchain network to obtain the node trust value update result.

[0037] Specifically, it addresses the lack of effective data integrity verification mechanisms and the difficulty in establishing trust relationships among participants in existing privacy protection schemes; it verifies signatures using asymmetric encryption algorithms to ensure the authenticity of participating nodes and the immutability of data; it utilizes blockchain consensus algorithms to confirm transaction validity, further enhancing the credibility of data records; and finally, by verifying the consistency between parameter update records in the distributed ledger and the original training process data, it generates credibility and integrity confirmation flags, updates node trust values, effectively establishes trust relationships among participants, ensures the transparency and reliability of the entire training process, and provides a solid foundation of trust for cross-institutional and cross-domain data collaboration.

[0038] S5. Automatically execute parameter verification rules through smart contracts, identify malicious node behavior based on preset anomaly detection thresholds, and trigger a penalty mechanism when parameter deviations exceed the normal range to obtain a reliable set of model parameters that have been verified by trust. Furthermore, a reliable set of model parameters that has been verified through trust is obtained, specifically including: S51. Obtain the parameter data submitted by the node through the smart contract, verify the parameters using the preset verification rules, and obtain a preliminary set of verified parameters. If there are parameters in the preliminary set of verified parameters that deviate from the normal range, they are marked as abnormal parameters to identify potential malicious nodes. S52. Based on the marked abnormal parameters, trigger the penalty mechanism in the smart contract to restrict the parameter submission permissions of malicious nodes, obtain the updated parameter set, and use the trust verification algorithm to perform secondary verification on the updated parameter set to obtain a trust score and determine the parameter set with high trust. S53. The high-confidence parameter set is aggregated through the federated learning algorithm to generate global model parameters. When the performance index of the global model parameters does not reach the preset threshold, the scoring criteria of the trust verification algorithm is adjusted, the secondary verification is re-executed, the optimized trust parameter set is obtained, the verification rules in the smart contract are updated, and the final reliable model parameter set is generated.

[0039] Specifically, by automatically executing parameter verification rules through smart contracts, malicious node behavior is identified and a penalty mechanism is triggered, ultimately obtaining a reliable set of model parameters that has been verified by trust. This approach effectively solves the problem of difficulty in identifying and preventing malicious node behavior in existing privacy protection schemes, ensuring the credibility and reliability of model parameters. Furthermore, by establishing an effective trust mechanism, the credibility and security of the model are improved, providing a reliable guarantee for cross-institutional data collaboration.

[0040] S6. Add calibration noise to the aggregated global model using differential privacy technology, control the noise intensity according to the privacy budget allocation strategy, and adjust the noise parameters when the privacy loss exceeds the preset upper limit to obtain the final collaborative model that meets the privacy protection standard. Furthermore, a final collaborative model that meets privacy protection standards is obtained, specifically including: The data distribution characteristics of the aggregated global model are obtained by differential privacy technology. The Laplace mechanism is used to add calibration noise to the model parameters to obtain an initial noise model. Then, according to the privacy budget allocation strategy, the noise intensity parameter is extracted from the initial noise model, and the noise intensity is adjusted by the exponential decay method to obtain an optimized noise model. When the privacy loss of the optimized noise model exceeds the preset threshold, the noise parameters are reconfigured through gradient pruning to obtain the adjusted noise model. Then, its differential privacy protection level is obtained, and the privacy loss is calculated through the ε-differential privacy formula (ε represents the privacy budget, ε=∑(Δf / σ), where Δf is the sensitivity of the query function and σ is the noise standard deviation). It is then determined whether the preset threshold is met to obtain the privacy compliant model. Model parameters are extracted from the privacy compliance model, and a federated learning framework is used for multi-party collaborative updates to obtain a collaborative update model. Then, the performance indicators of the global model are obtained through the aggregated parameters of the collaborative update model to determine whether the privacy protection standards are met, thus obtaining the final collaborative model.

[0041] Specifically, differential privacy technology is used to add calibration noise to the aggregated global model, and the noise intensity is controlled according to the privacy budget allocation strategy. Finally, a collaborative model that meets the privacy protection standard is obtained. This method effectively solves the problem of insufficient privacy protection in the model aggregation process and ensures that the global model has good performance while meeting the privacy protection requirements.

[0042] S7. Perform quality checks on the collaborative training results based on model performance evaluation metrics. Test the model's generalization ability on data at each node using cross-validation. When the accuracy reaches the expected target, the collaborative modeling process is completed, and a deployable privacy-preserving joint learning model is determined.

[0043] Furthermore, deployable privacy-preserving joint learning models are identified, specifically including: The training and validation subsets are obtained from the data of each node through cross-validation to obtain the partitioned dataset. The collaborative model is trained using the K-fold cross-validation method to obtain the prediction results of each fold. The performance metrics, including accuracy and loss value, are obtained from the prediction results to determine the model's generalization ability on data at each node. When the accuracy reaches a preset threshold, the model parameters of each node are aggregated through a joint learning framework to obtain the global model. A differential privacy algorithm is used to add noise to obtain a privacy-preserving model. The performance metrics of the privacy-preserving model are tested using a validation set to determine whether it meets the deployability requirements. If the performance metrics meet a preset threshold, a deployable privacy-preserving joint learning model is determined.

[0044] Specifically, the collaborative training results are evaluated for quality using model performance metrics, and cross-validation is used to test the model's generalization ability on data from each node. Ultimately, a deployable privacy-preserving joint learning model is determined, effectively solving the problem of difficulty in evaluating model generalization ability and ensuring model performance during joint modeling, and ensuring the stability and reliability of the model on different datasets.

[0045] This invention forms a complete solution from data encryption, privacy protection, distributed training, model aggregation, trust verification to final model evaluation and deployment, effectively solving key problems such as data silos, insufficient privacy protection, and lack of trust mechanisms. Through homomorphic encryption and differential privacy technologies, it ensures the privacy and security of data during sharing and use. Utilizing federated learning and secure multi-party computation, it achieves efficient aggregation of distributed training and model parameters. Leveraging blockchain technology, it establishes trust relationships among participants, ensuring the credibility and integrity of the training process. Finally, through model performance evaluation and cross-validation, it determines a deployable privacy-preserving federated learning model, providing a high-performance, high-reliability solution for cross-institutional and cross-domain data collaboration, promoting the full release of big data value, and meeting stringent data protection regulations and privacy requirements.

[0046] Example 2

[0047] Please see Figure 4 This embodiment provides a big data privacy protection modeling system based on federated learning and blockchain, used to implement a big data privacy protection modeling method based on federated learning and blockchain, including: The data encryption and dimensionality reduction module performs homomorphic encryption and dimensionality reduction on the original data of the participating nodes to generate encrypted training samples. The federated learning training module trains a local model on an encrypted dataset using a federated learning framework, and uses the gradient descent algorithm to calculate the weights of the local model, thus enabling the training and updating of the local model. The multi-party computation and model aggregation module uses a secure multi-party computation protocol to aggregate encrypted model parameters from each node, and calculates global model parameters through a secret sharing mechanism and predefined weights to generate a joint model. The verification and trust confirmation module constructs a blockchain verification network to perform hash identification and signature verification on model parameters and training process, and writes parameter update records into a distributed ledger. The smart contract and trust verification module automatically executes parameter verification rules through smart contracts, identifies malicious node behavior and triggers a penalty mechanism, performs secondary verification of the parameter set, and generates a reliable model parameter set. The differential privacy model generation module adds calibration noise to the aggregated global model, controls the noise intensity according to the privacy budget, adjusts the noise parameters, and generates a privacy-preserving collaborative model. The model evaluation and deployment module performs quality checks on the collaborative training results based on performance evaluation metrics, tests the model's generalization ability through cross-validation, and determines deployable privacy-preserving joint learning models.

[0048] Example 3

[0049] Based on Example 1, this embodiment further introduces a dynamic privacy budget allocation mechanism and an adaptive differential privacy parameter adjustment algorithm to enhance the model's ability to dynamically balance privacy protection and performance. By monitoring privacy risks and performance indicators during the model training process in real time, the privacy budget allocation strategy is dynamically adjusted and the differential privacy parameters are optimized, thereby maximizing the model's performance while protecting data privacy.

[0050] In practical applications, the sensitivity and privacy risks of data change dynamically during the training process. Therefore, this embodiment introduces a dynamic privacy budget allocation mechanism.

[0051] Specifically, at the beginning of each iteration of model training, the privacy risk assessment module monitors the privacy risk level of the current dataset in real time. The privacy risk assessment module dynamically calculates the privacy risk value based on the sensitivity of the data, changes in the data distribution, and the current state of the model. Based on the privacy risk value, the privacy budget allocation is dynamically adjusted. When the privacy risk is high, the privacy budget allocation is increased to enhance privacy protection; when the privacy risk is low, the privacy budget allocation is appropriately reduced to optimize model performance. The adjusted privacy budget is allocated to various privacy protection aspects, including homomorphic encryption and differential privacy noise addition, to ensure a dynamic balance between privacy protection and performance throughout the training process.

[0052] The addition of differential privacy noise is based on preset fixed parameters. However, fixed parameters are difficult to adapt to changes in different training stages and data distributions. Therefore, this embodiment introduces an adaptive differential privacy parameter adjustment algorithm.

[0053] Specifically, in each iteration cycle, the performance monitoring module monitors the model's performance metrics, such as accuracy and loss value, in real time. Based on the trends in these performance metrics, the parameters of the differential privacy noise are adaptively adjusted. For example, when model performance degrades significantly, the noise intensity is appropriately reduced; when model performance is stable and privacy risk is low, the noise intensity is appropriately increased. Through a feedback mechanism, the adjusted noise parameters are applied to the differential privacy noise addition process in the next iteration cycle, ensuring that the model maximizes performance while protecting privacy.

[0054] Dynamic privacy budget allocation mechanism: By monitoring privacy risks in real time and dynamically adjusting the privacy budget allocation, the balance between privacy protection and performance is made more flexible and efficient. This mechanism can dynamically adjust the strength of privacy protection based on actual data and model status, avoiding the limitations of static privacy budget allocation.

[0055] Among them, the adaptive differential privacy parameter adjustment algorithm: by monitoring the model performance in real time, it adaptively adjusts the differential privacy noise parameter, so that the model can maintain the best balance between privacy protection and performance under different training stages and data distributions. This algorithm can dynamically adapt to changes in the model training process, improving the robustness and practicality of the model.

[0056] By introducing a dynamic privacy budget allocation mechanism and an adaptive differential privacy parameter adjustment algorithm, this embodiment significantly improves the performance and adaptability of the model while protecting data privacy, providing a more efficient and reliable solution for cross-institutional and cross-domain big data privacy protection modeling.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A big data privacy protection modeling method based on federated learning and blockchain, characterized by: Includes the following steps: The original training data of each participating node is encrypted using a homomorphic encryption algorithm. An encrypted dataset is generated based on a preset encryption key and then dimensionality reduction is performed to obtain encrypted training samples that meet privacy protection requirements. The local model parameters are trained on each encrypted dataset using a federated learning framework. The local model weight vector is calculated using the gradient descent algorithm to obtain the encrypted model parameters to be aggregated. A secure multi-party computation protocol is used to aggregate the encrypted model parameters of each node. The parameters are then transmitted to other collaborating nodes in fragments through a secret sharing mechanism. The global model parameters are calculated based on the predefined aggregation weights to obtain a joint model that integrates the features of multi-party data. Obtain the digital signature information of each participating node and build a blockchain verification network. Generate unique identifiers for model parameters and training process based on hash algorithm to determine the credibility and integrity of the training process. The parameter verification rules are automatically executed by smart contracts to identify malicious node behavior based on preset anomaly detection thresholds, thereby obtaining a reliable set of model parameters that have been verified by trust. By adding calibration noise to the aggregated global model using differential privacy technology, and controlling the noise intensity according to the privacy budget allocation strategy, a final collaborative model that meets privacy protection standards is obtained. The collaborative training results are quality-checked based on model performance evaluation metrics. The generalization ability of the model on data at each node is tested using cross-validation methods to determine deployable privacy-preserving joint learning models.

2. The big data privacy protection modeling method based on federated learning and blockchain according to claim 1, characterized in that: Obtain encrypted training samples that meet privacy protection requirements, specifically including: The original training data of each participating node is encrypted using a homomorphic encryption algorithm. An encrypted dataset is generated based on a preset encryption key. Principal component analysis is then performed on the encrypted dataset to reduce its dimensionality, thereby obtaining the encrypted feature set after dimensionality reduction and determining the low-dimensional encrypted feature set. When the dimension of the low-dimensional encrypted feature set meets the preset threshold, a clustering algorithm is used to group the low-dimensional encrypted feature set to obtain encrypted sample groups. When the dimension does not meet the preset threshold, the principal component analysis parameters are adjusted and the low-dimensional encrypted feature set is obtained again. Based on the encrypted sample grouping, noise is added using a differential privacy algorithm to obtain the encrypted training samples with added noise, and the encrypted training samples that meet the privacy protection requirements are determined. Feature distribution information is extracted from the encrypted training samples that meet the privacy protection requirements, and it is judged whether the feature distribution information conforms to the preset statistical characteristics to obtain the feature distribution verification result. If the feature distribution verification result meets the preset statistical characteristics, then the metadata description of the encrypted training sample is generated based on the feature distribution information, and the metadata of the encrypted training sample is obtained. If it does not meet the requirements, then the noise parameters of the differential privacy algorithm are adjusted, and the encrypted training sample is obtained again. Based on the encrypted training sample metadata, a data verification tool is used to verify the integrity and consistency of the metadata, thus obtaining the final encrypted training sample that meets privacy protection requirements.

3. The big data privacy protection modeling method based on federated learning and blockchain according to claim 2, characterized in that: Data validation tools are used to verify the integrity and consistency of metadata, specifically including: A consistency check algorithm is used to verify the metadata description of the encrypted training samples and obtain the consistency check result. Then, an integrity check algorithm is used to check the integrity of the metadata description and obtain the integrity check result. Based on the integrity verification results, a feature extraction template is generated using the metadata description of the encrypted training samples. The feature extraction template is then obtained, and features are extracted from the encrypted training samples to obtain the extracted feature set. If the distribution characteristics of the feature set meet the preset statistical criteria, a classification algorithm is used to group the feature set and obtain the encrypted sample set after grouping. If it does not meet the criteria, the feature extraction template parameters are adjusted and the feature set is obtained again. Based on the grouped encrypted sample set, generate metadata tags for the sample set, obtain the encrypted sample tag set, and verify the encrypted sample tag set through a consistency check algorithm.

4. The big data privacy protection modeling method based on federated learning and blockchain according to claim 1, characterized in that: Obtain the parameters of the encryption model to be aggregated, specifically including: The federated learning framework obtains training data from various encrypted datasets, updates the weight vector of the local model using the gradient descent algorithm, and obtains the local model weights. When the training loss function value of the local model weights is lower than the preset convergence threshold, the local training is considered complete, and the encrypted model parameters are obtained. Based on the encrypted model parameters, a secure multi-party computation protocol is used to aggregate the parameters of each local model to obtain the global model parameters, update the weight vector of each local model, and obtain the updated local model. When the loss function value of the updated local model is lower than the global convergence threshold, the global model is determined to have converged, and the final global model is obtained. Then, differential privacy technology is used to perturb the model parameters to obtain privacy-preserving global model parameters, which are then distributed to each local node to update the local model and obtain the final encrypted model parameters to be aggregated.

5. The big data privacy protection modeling method based on federated learning and blockchain according to claim 1, characterized in that: The resulting joint model, which integrates features from multiple data sources, includes: A secure multi-party computation protocol is adopted to obtain encrypted model parameters from each node, perform parameter sharding to obtain a sharded parameter set, and transmit the sharded parameter set to the cooperating node network through a secret sharing mechanism to ensure that each node receives the complete sharded parameters. Once the receiving node verifies the integrity of the sharding parameters, it performs a weighted aggregation operation based on the predefined aggregation weights to obtain preliminary global model parameters. Then, it adds noise using a differential privacy algorithm to obtain privacy-enhanced global model parameters. Through a distributed computing architecture, privacy-enhancing global model parameters are distributed to each collaborating node, and the updated local model parameters of each node are obtained. When the deviation between the local model parameters and the global model parameters is less than a preset threshold, parameter fusion is performed to obtain a joint model that integrates the features of multiple data sources.

6. The big data privacy protection modeling method based on federated learning and blockchain according to claim 1, characterized in that: Determining the reliability and integrity of the training process specifically includes: Obtain the digital signatures of each participating node, verify the signatures using an asymmetric encryption algorithm, and obtain the signature verification result. If the signature verification result passes, extract the model parameters and training process data, generate a unique identifier using the SHA-256 hash algorithm, and obtain the unique identifier value. Based on the unique identifier value, the transaction records in the blockchain verification network are constructed, and the validity of the transaction is confirmed by the consensus algorithm to obtain the transaction confirmation status. When the transaction confirmation status is valid, the model parameter update record is written into the distributed ledger to obtain the ledger update status. Based on the ledger update status, the consistency between the parameter update records in the distributed ledger and the original training process data is verified. When the consistency verification result passes, a trustworthiness confirmation flag and an integrity confirmation flag are generated to obtain the trustworthiness and integrity status of the training process. These are used to update the node trust value in the blockchain network to obtain the node trust value update result.

7. The big data privacy protection modeling method based on federated learning and blockchain according to claim 1, characterized in that: Obtain a reliable set of model parameters that has been verified by trust, specifically including: The system obtains parameter data submitted by nodes through smart contracts, verifies the parameters using preset verification rules, and obtains a preliminary set of verified parameters. If there are parameters in the preliminary set of verified parameters that deviate from the normal range, they are marked as abnormal parameters, thus identifying potential malicious nodes. Based on the marked abnormal parameters, the penalty mechanism in the smart contract is triggered to restrict the parameter submission permissions of malicious nodes, resulting in an updated parameter set. The updated parameter set is then verified a second time using a trust verification algorithm to obtain a trust score and determine the parameter set with high trust. The federated learning algorithm aggregates the set of high-trust parameters to generate global model parameters. When the performance index of the global model parameters does not reach the preset threshold, the scoring criteria of the trust verification algorithm is adjusted, and a second verification is performed to obtain an optimized set of trust parameters. The verification rules in the smart contract are then updated to generate the final set of reliable model parameters.

8. The big data privacy protection modeling method based on federated learning and blockchain according to claim 1, characterized in that: The final collaborative model that meets privacy protection standards includes: The data distribution characteristics of the aggregated global model are obtained by differential privacy technology. The Laplace mechanism is used to add calibration noise to the model parameters to obtain an initial noise model. Then, according to the privacy budget allocation strategy, the noise intensity parameter is extracted from the initial noise model, and the noise intensity is adjusted by the exponential decay method to obtain an optimized noise model. When the privacy loss of the optimized noise model exceeds the preset threshold, the noise parameters are reconfigured through gradient pruning to obtain the adjusted noise model. Then, its differential privacy protection level is obtained, and the privacy loss is calculated through the ε-differential privacy formula to determine whether the preset threshold is met, thus obtaining a privacy-compliant model. Model parameters are extracted from the privacy compliance model, and a federated learning framework is used for multi-party collaborative updates to obtain a collaborative update model. Then, the performance indicators of the global model are obtained through the aggregated parameters of the collaborative update model to determine whether the privacy protection standards are met, thus obtaining the final collaborative model.

9. The big data privacy protection modeling method based on federated learning and blockchain according to claim 1, characterized in that: Identify deployable, privacy-preserving joint learning models, specifically including: The training and validation subsets are obtained from the data of each node through cross-validation to obtain the partitioned dataset. The collaborative model is trained using the K-fold cross-validation method to obtain the prediction results of each fold. The performance metrics, including accuracy and loss value, are obtained from the prediction results to determine the model's generalization ability on data at each node. When the accuracy reaches a preset threshold, the model parameters of each node are aggregated through a joint learning framework to obtain the global model. A differential privacy algorithm is used to add noise to obtain a privacy-preserving model. The performance metrics of the privacy-preserving model are tested using a validation set to determine whether it meets the deployability requirements. If the performance metrics meet a preset threshold, a deployable privacy-preserving joint learning model is determined.

10. A big data privacy-preserving modeling system based on federated learning and blockchain, used to implement the big data privacy-preserving modeling method based on federated learning and blockchain as described in any one of claims 1-9, characterized in that: include: The data encryption and dimensionality reduction module performs homomorphic encryption and dimensionality reduction on the original data of the participating nodes to generate encrypted training samples. The federated learning training module trains a local model on an encrypted dataset using a federated learning framework, and uses the gradient descent algorithm to calculate the weights of the local model, thus enabling the training and updating of the local model. The multi-party computation and model aggregation module uses a secure multi-party computation protocol to aggregate encrypted model parameters from each node, and calculates global model parameters through a secret sharing mechanism and predefined weights to generate a joint model. The verification and trust confirmation module constructs a blockchain verification network to perform hash identification and signature verification on model parameters and training process, and writes parameter update records into a distributed ledger. The smart contract and trust verification module automatically executes parameter verification rules through smart contracts, identifies malicious node behavior and triggers a penalty mechanism, performs secondary verification of the parameter set, and generates a reliable model parameter set. The differential privacy model generation module adds calibration noise to the aggregated global model, controls the noise intensity according to the privacy budget, adjusts the noise parameters, and generates a privacy-preserving collaborative model. The model evaluation and deployment module performs quality checks on the collaborative training results based on performance evaluation metrics, tests the model's generalization ability through cross-validation, and determines deployable privacy-preserving joint learning models.

Images