Method and device for verifying high-dimensional heterogeneous parameters of federated learning model

Abstract

The invention relates to the technical field of federated learning and block chain crossing, in particular to a verification method and device for high-dimensional heterogeneous parameters of a federated learning model. The method comprises the steps that an intelligent contract selects an institution meeting the trust degree requirement from all institutions participating in multi-centralized federated learning as a verification node; uploading the high-dimensional heterogeneous parameters of the local model to a block chain by each mechanism, and triggering an intelligent contract verification process; the verification nodes verify the high-dimensional heterogeneous parameters of all the local models, the intelligent contract collects initial verification results of all the verification nodes, and a final consensus verification result is determined through a consensus mechanism; and the smart contract writes the final consensus verification result, the parameter details and the verification record of each node into the block chain for evidence storage, authorizes the parameters passing the verification to participate in global model aggregation, and rejects the parameters not passing the verification. The problems that federal learning depends on a third party to perform model parameter aggregation, model parameter verification lacks a decentration mechanism and a verification result does not have a tampering-free evidence can be solved.

Description

Technical Field

[0001] This invention relates to the field of federated learning and blockchain technology, and in particular to a method and apparatus for verifying high-dimensional heterogeneous parameters of a federated learning model. Background Technology

[0002] With the rapid development of the digital economy, the demand for cross-institutional data collaboration is becoming increasingly urgent, and data privacy protection has become a core requirement. Federated learning, with its characteristic of "data usable but not visible," has become a key technology for cross-domain data value mining and is widely used in collaborative fields requiring privacy protection, such as finance, healthcare, and credit reporting.

[0003] Existing federated learning systems generate a global model by aggregating local model parameters from participating institutions. Some solutions attempt to incorporate blockchain technology to leverage its decentralized nature and facilitate collaboration, but deep integration between the two has not been achieved. While blockchain smart contracts possess automatic execution capabilities, they have not been effectively applied to collaborative model parameter verification scenarios.

[0004] Existing technologies have significant drawbacks: federated learning relies on third parties to aggregate model parameters, which poses a single point of failure and the risk of model contamination; model parameter verification lacks a decentralized mechanism, making it difficult to guarantee authenticity and susceptible to data discrimination or model poisoning attacks; verification results lack tamper-proof evidence, making auditing and traceability difficult, and there is a lack of robust verification schemes for high-dimensional heterogeneous parameters. Summary of the Invention

[0005] This invention provides a method and apparatus for verifying high-dimensional heterogeneous parameters of a federated learning model, in order to solve the problems in the prior art where federated learning relies on third parties for model parameter aggregation, model parameter verification lacks a decentralized mechanism, and verification results lack tamper-proof evidence.

[0006] In a first aspect, embodiments of the present invention provide a method for verifying high-dimensional heterogeneous parameters of a federated learning model, comprising: The smart contract selects institutions that meet the trust requirements from among the participating institutions in the multi-centralized federated learning process to form a set of verification nodes. After each institution completes the training of its local model, it uploads the high-dimensional heterogeneous parameters of the local model to the blockchain, triggering the smart contract verification process. The verification node verifies all the high-dimensional heterogeneous parameters of the uploaded local model. The smart contract collects the initial verification results of each verification node and determines the final consensus verification result through a multi-node consensus mechanism. The smart contract writes the final consensus verification result, parameter details, and verification records of each node into the blockchain for storage. Parameters that pass the verification are authorized to participate in the global model aggregation, while parameters that fail the verification are rejected.

[0007] In one possible implementation, the smart contract selects institutions that meet the trust requirements from among the participating institutions in the multi-centralized federated learning as verification nodes, including: The smart contract obtains the historical verification accuracy, computing power contribution value, and compliance score of each institution participating in the multi-centralized federated learning; The Improved Practical Byzantine Fault Tolerance (PBFT) algorithm is used to calculate the trust value of each institution based on the historical verification accuracy, computing power contribution value, and compliance score. When the trust value is greater than or equal to the trust threshold, the institution corresponding to the trust value is determined as a verification node.

[0008] In one possible implementation, an improved PBFT algorithm is used to calculate the trust score of each institution based on the historical verification accuracy, computing power contribution value, and compliance score, including: according to Calculate the trust score for each institution; in, Indicates the first Trust rating of an organization , , These are the weighting coefficients, and , Indicates the first Historical verification accuracy of individual institutions Indicates the first The computing power contribution of each institution. Indicates the first Compliance scores for each organization.

[0009] In one possible implementation, the verification node verifies all uploaded local model high-dimensional heterogeneous parameters, including: The verification node sequentially performs parameter anomaly dimension pre-screening, parameter standardization and robust conversion, multidimensional consistency check and fine-grained detection on all uploaded local model high-dimensional heterogeneous parameters.

[0010] In one possible implementation, the parameter anomaly dimension pre-screening is performed by using an improved parallel isolated forest algorithm to calculate anomaly scores for each dimension of the high-dimensional heterogeneous parameters of the local model. When the anomaly score is greater than a preset anomaly threshold, the current dimension is marked as an anomaly dimension and deleted. The improved parallel isolated forest algorithm introduces time decay weights and a sparse adaptive sampling mechanism on the basis of the existing isolated forest algorithm. The calculation formula for the improved parallel isolated forest algorithm is as follows: ; in, Indicates the first Outlier scores in dimensions Indicates the first The first organization uploaded The set of parameters for the dimension. Indicates the time decay weight. , This represents the difference between the timestamp of the current parameter upload and the system's base time. Indicates the preset time decay period. Indicates a sparse adaptive sampling rate. , Indicates the first The first organization uploaded The sum of the absolute values ​​of all parameters of the dimension. This represents the maximum value of the sum of the absolute values ​​of all dimension parameters.

[0011] In one possible implementation, the parameter standardization and robustness are converted into standardization of the parameters of the remaining dimensions after pre-screening for abnormal parameters using the median and the absolute deviation of the median. The formula for standardization is: ; in, Indicates the first The first organization uploaded Standardized parameters of dimensional parameters, Indicates the first The first organization uploaded Dimensional parameters, Indicates the first The first organization uploaded The median of the parameter set of the dimension. Indicates the first The first organization uploaded The median absolute deviation of the parameters of the dimension.

[0012] In one possible implementation, the multidimensional consistency test is to use an improved Mahalanobis distance that introduces regularized covariance estimation and feature selection mechanism to calculate the Mahalanobis distance between the standardized parameter and the global parameter, and when the Mahalanobis distance is greater than a preset dynamic threshold, the current standardized parameter is determined to be an abnormal parameter and marked. The improved formula for calculating Mahalanobis distance is: ; in, Indicates the first Mahalanobis distance between the standardized parameters of each institution and the global parameters. Indicates the first Standardized parameters for each organization This represents the global parameter mean vector. This represents the global parameter covariance matrix after feature filtering. Represents the regularization coefficient. Represents the identity matrix.

[0013] In one possible implementation, the fine-grained detection employs an optimized local anomaly factor algorithm that incorporates dynamic neighborhood selection and weighted distance measurement mechanisms. This algorithm calculates a local anomaly factor on the standardized parameters after multidimensional consistency testing, and determines that the current standardized parameter is anomaly in the local neighborhood when the local anomaly factor is greater than a preset anomaly factor threshold. The calculation formula for the optimized local anomaly factor algorithm is as follows: ; in, Indicates the first Local outlier factors after optimization of standardized parameters for each institution. Indicates the first Standardized parameters for each organization nearest neighbor set Indicates the first Standardized parameters for each organization The number of elements in the nearest neighbor set. Indicates distance weight, Represents the nearest neighbor set of the i-th Standardized parameter vectors for each institution, Represents the neighborhood scale parameter. Represents the nearest neighbor set of the i-th The locally achievable density of an individual institution. Indicates the first The local reachability density of standardized parameters for an organization.

[0014] In one possible implementation, the final consensus verification result is determined through a multi-node consensus mechanism, including: according to Determine the final consensus verification result; in, Indicates the first The final consensus verification results of the standardized parameters of each institution. Indicates the first The similarity weight of each verification node. , This represents the total number of verification nodes. Represents the similarity enhancement coefficient. Indicates the first The consistency of the verification results of each verification node with the verification results of other verification nodes. This represents the historical verification reliability coefficient of the verification node. Indicates the first The verification node verifies the first... Initial validation results of standardized parameters for each organization. This represents the contribution adjustment coefficient. Indicates the first The contribution weight of each verification node.

[0015] Secondly, embodiments of the present invention provide a verification apparatus for high-dimensional heterogeneous parameters of a federated learning model, comprising: The smart contract module is used to select institutions that meet the trust requirements from among the participating institutions in the multi-centralized federated learning as verification nodes, forming a set of verification nodes. Each institution's terminal is used to upload the high-dimensional heterogeneous parameters of the local model to the blockchain after completing the local model training, triggering the smart contract verification process; The verification node terminal is used to verify the high-dimensional heterogeneous parameters of all uploaded local models and send the initial verification results to the smart contract module. The smart contract module is also used to collect the initial verification results of each verification node, determine the final consensus verification result through a multi-node consensus mechanism, and write the final consensus verification result, parameter details and verification records of each node into the blockchain for storage. Parameters that pass the verification are authorized to participate in the global model aggregation, while parameters that fail the verification are rejected.

[0016] This invention provides a method and apparatus for verifying high-dimensional heterogeneous parameters of a federated learning model. Through smart contracts, institutions meeting trust requirements are selected from among the participating institutions in a multi-centralized federated learning system to serve as verification nodes, forming a verification node set. After each institution completes local model training, it uploads the high-dimensional heterogeneous parameters of that local model to the blockchain, triggering the smart contract verification process. Verification nodes verify all uploaded high-dimensional heterogeneous parameters of the local model. The smart contract collects the initial verification results from each verification node and determines the final consensus verification result through a multi-node consensus mechanism. The smart contract writes the final consensus verification result, parameter details, and verification records of each node into the blockchain for notarization. Parameters that pass verification are authorized to participate in global model aggregation, while parameters that fail verification are rejected. This invention, by selecting institutions meeting trust requirements as verification nodes through smart contracts for subsequent verification of high-dimensional heterogeneous parameters of the local model, replaces the traditional third-party parameter aggregation model, completely resolving the risk of single-point failures and avoiding global model pollution caused by dishonest third-party behavior, thereby improving the stability and security of the federated learning system.

[0017] This invention verifies all the high-dimensional heterogeneous parameters of the uploaded local models through verification nodes, ensuring that the nodes participating in the verification have high credibility; combined with a multi-node consensus mechanism to aggregate the initial verification results, it effectively resists malicious attacks such as data discrimination and model poisoning, and ensures the authenticity and validity of the local model parameters.

[0018] This invention uses smart contracts to completely write the final consensus verification result, parameter details, and verification records of each node into the blockchain for evidence storage, forming a full-process audit chain. This solves the problems of traditional verification results being difficult to trace and responsibility being unclear, and provides a reliable compliance basis for cross-institutional collaboration. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the implementation of the verification method for high-dimensional heterogeneous parameters of the federated learning model provided in this embodiment of the invention. Figure 2 This is a schematic diagram of the structure of the verification device for high-dimensional heterogeneous parameters of the federated learning model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the terminal provided in an embodiment of the present invention. Detailed Implementation

[0021] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, specific embodiments will be described below in conjunction with the accompanying drawings.

[0023] Figure 1 The implementation flowchart of a method for verifying high-dimensional heterogeneous parameters of a federated learning model provided in this embodiment of the invention is detailed below: Step 101: The smart contract selects institutions that meet the trust requirements from among the participating institutions in the multi-centralized federated learning as verification nodes, thus forming a set of verification nodes.

[0024] The core purpose of this step is to achieve decentralized selection of verification nodes through smart contracts, select highly trustworthy institutions to participate in subsequent model parameter verification, lay a safe and reliable node foundation for the entire multi-centralized federated learning model parameter verification process, and avoid verification failure caused by the participation of a single node or an untrustworthy node.

[0025] In this step, the smart contract first invokes the blockchain-modified PBFT algorithm embedded within it. This algorithm provides core technical support for the security classification and selection of verification nodes, adapting to the decentralized collaboration needs in cross-institutional collaboration scenarios and ensuring the fairness and attack resistance of the selection process. Based on this algorithm, the smart contract will perform a quantitative assessment of the trust level of all institutions participating in the multi-centralized federated learning, selecting only those institutions whose trust level meets the preset threshold as verification nodes. Only after a set of multi-centralized verification nodes meeting the trust level requirement is determined can subsequent operations such as parameter anomaly detection, consistency verification, and consensus voting be collaboratively executed by the verification nodes. Institutions that fail the trust level screening cannot participate in the verification work.

[0026] In one embodiment, the smart contract selects an institution that meets the trust requirements as a verification node from among the institutions participating in the multi-centralized federated learning, which may include: The smart contract obtains the historical verification accuracy, computing power contribution value, and compliance score of each institution participating in the multi-centralized federated learning; An improved PBFT algorithm is used to calculate the trust score of each institution based on historical verification accuracy, computing power contribution value, and compliance score. When the trust value is greater than or equal to the trust threshold, the institution corresponding to the trust value is determined as the verification node.

[0027] Optionally, an improved PBFT algorithm is used to calculate the trust score for each institution based on historical verification accuracy, computing power contribution, and compliance score, including: according to Calculate the trust score for each institution; in, Indicates the first Trust rating of an organization , , These are the weighting coefficients, and , , , The values ​​of the three weighting coefficients are derived from multi-scenario testing, ensuring the scientific and reasonable nature of trust assessment. This is further supported by the core requirements of federated learning for cross-institutional collaboration scenarios. This is the weighting coefficient corresponding to the historical verification accuracy rate, which can be 0.4. Since the accuracy of the institution's past verifications directly determines the reliability of the current verification results, it is a core consideration. The weighting coefficient for the contribution value of computing power can be 0.3. Computing power support is the basic guarantee for the efficient advancement of the verification process. This is the compliance scoring weighting coefficient, which can take a value of 0.3. Compliance is a prerequisite for cross-organizational collaboration. Indicates the first The historical verification accuracy rate of an organization is equal to the number of correct verifications divided by the total number of verifications. Indicates the first The computing power contribution value of each institution is equal to that of the first institution. The verification computing power of each institution divided by the average verification computing power of all institutions. Indicates the first The compliance score of each organization is given by the regulatory node based on compliance records (such as data processing compliance, collaboration process compliance, etc.), and the value range is [0, 1].

[0028] Optional, The calculation process is as follows: The weight of data processing compliance is set at 0.5, which includes the legality of data collection, the security of data storage, and the compliance of data transmission. Among them, the legality of data collection refers to whether the data collection has been authorized by the user, accounting for 0.2; the security of data storage refers to whether the data storage complies with the encryption storage standard, accounting for 0.15; and the compliance of data transmission refers to whether the data transmission is carried out through a secure transmission protocol, accounting for 0.15. The weight of compliance of the collaboration process is set to 0.5, which includes the completeness of the collaboration agreement signing, the standardization of the task execution process, and the timeliness of reporting abnormal situations. Among them, the completeness of the collaboration agreement signing can be set to 0.2, the standardization of the task execution process to 0.2, and the timeliness of reporting abnormal situations to 0.1.

[0029] Regulatory nodes score each of the above-mentioned sub-indicators (each indicator has a maximum score of 1 point), and then calculate the result by weighting the scores according to their respective weights. .

[0030] The trust threshold can be set in the range of [0.8, 0.9]. This trust threshold setting ensures the high credibility of the verification nodes while avoiding an insufficient number of selectable verification nodes due to an excessively high trust threshold.

[0031] Ultimately, all institutions that meet the trust requirements will collectively form a set of verification nodes. The number of verification nodes in the set Must meet Furthermore, the number is odd. This number setting can adapt to the needs of subsequent multi-node consensus verification mechanisms, ensuring the uniqueness and validity of consensus results, and providing structural support for consensus decisions in subsequent parameter verification.

[0032] Step 102: After each institution completes the training of its local model, it uploads the high-dimensional heterogeneous parameters of the local model to the blockchain, triggering the smart contract verification process.

[0033] Each participating institution must first complete the training process of its own local model. After training, each institution needs to extract and organize the high-dimensional heterogeneous parameters of its local model. The high-dimensional heterogeneous parameters of the local model are represented in set form as follows: , Used to identify institutions participating in federal learning The parameter dimension represents the high-dimensional heterogeneous parameters of the local model, because federated learning models typically have complex structures. The values ​​of are usually large, reflecting the "high-dimensional" characteristics of the parameter; however, the sources, distributions, and training strategies of training data may differ among institutions, leading to variations in the data uploaded by different institutions. There are differences in parameter distribution and numerical range, i.e., the "heterogeneous" characteristics of the parameters, which is also the core scenario that the subsequent verification process needs to adapt to.

[0034] After each institution completes the above parameter compilation, it will convert the high-dimensional heterogeneous parameters of the local model. All parameters are uploaded to the blockchain network. The decentralized storage characteristic of the blockchain ensures the transparency and immutability of the uploading process of high-dimensional heterogeneous parameters of the local model, preventing the high-dimensional heterogeneous parameters of the local model from being tampered with or replaced during transmission or storage, and ensuring the originality of the high-dimensional heterogeneous parameters of the local model to be verified.

[0035] When all participating institutions have completed uploading the high-dimensional heterogeneous parameters of their local models, or when the deadline for uploading these parameters, as preset by the smart contract, is met, the blockchain network will automatically trigger the smart contract verification process embedded within it. The upload deadline only requires meeting any of the following conditions: (1) Time deadline: The smart contract has a preset fixed upload duration. The timer starts from the timestamp of the first institution uploading parameters and reaches [a certain point]. This triggers the verification process; The value can be adjusted according to the collaboration scenario; the default value is 24 hours. (2) Quantity cutoff condition: Preset threshold for the proportion of participating institutions that upload data. (Value range [0.8, 1.0], default 0.9), when the proportion of the number of organizations that have uploaded parameters to the total number of participating organizations reaches... When this occurs, the verification process is triggered.

[0036] This triggering mechanism requires no human intervention and relies entirely on the automatic execution capability of smart contracts to ensure the timeliness and fairness of the verification process.

[0037] After the smart contract triggers the verification process, it will follow the preset logic to call subsequent algorithms and rules for parameter validity screening, multi-node consensus verification, etc., and formally enter the verification stage of high-dimensional heterogeneous parameters of the local model. The core verification stage proceeds step by step in the order of "anomaly dimension pre-screening → parameter standardization and robustness conversion → multi-dimensional consistency verification". It is a progressive process of parameter "coarse screening → fine processing → fine verification", as described in step 103.

[0038] Step 103: The verification nodes verify all the high-dimensional heterogeneous parameters of the uploaded local models. The smart contract collects the initial verification results of each verification node and determines the final consensus verification result through a multi-node consensus mechanism.

[0039] This step is the core execution link of the entire parameter verification process. Its core purpose is to perform multi-stage and multi-level validity screening of the uploaded local model high-dimensional heterogeneous parameters through verification nodes. Then, the smart contract aggregates the initial verification results through a multi-node consensus mechanism to ensure the accuracy, robustness and anti-attack of the final verification results, and to provide a decision basis for whether subsequent parameters should participate in the global model aggregation.

[0040] In one embodiment, the verification node verifies all uploaded local model high-dimensional heterogeneous parameters, including: The verification node sequentially performs parameter anomaly dimension pre-screening, parameter standardization and robust conversion, multidimensional consistency check and fine-grained detection on all uploaded local model high-dimensional heterogeneous parameters.

[0041] Optionally, pre-screening of abnormal parameter dimensions is the first step. The purpose is to remove obviously abnormal parameter dimensions, reduce the amount of data to be processed later, and avoid interference from invalid dimensions on the verification results. The set of valid dimensions output is the basic data range for subsequent standardization and consistency testing.

[0042] The parameter anomaly dimension pre-screening adopts an improved parallel isolated forest algorithm to calculate the anomaly score of each dimension of the high-dimensional heterogeneous parameters of the local model. When the anomaly score is greater than the preset anomaly threshold, the current dimension is marked as an anomaly dimension and deleted. The parameters corresponding to the anomaly dimension are not needed when the smart contract determines the final consensus verification result through the multi-node consensus mechanism, so they are directly deleted.

[0043] The improved parallel isolated forest algorithm introduces time decay weights and a sparse adaptive sampling mechanism on the basis of the existing isolated forest algorithm, which effectively improves the anomaly identification efficiency of high-dimensional sparse parameters. The calculation formula for the improved parallel isolated forest algorithm is as follows: ; in, Indicates the first The anomaly score for a dimension ranges from [0, 1]. A score closer to 1 indicates a higher probability of an anomaly in that dimension. Indicates the first The first organization uploaded The set of parameters for the dimension. , This indicates the total number of participating organizations. Indicates the time decay weight. , It can be adjusted according to the parameter update frequency. This represents the difference between the timestamp of the current parameter upload and the system's base time. Indicates the preset time decay period. Used to reduce the interference of outdated parameters on the verification results. Indicates a sparse adaptive sampling rate. , Indicates the first The first organization uploaded The sum of the absolute values ​​of all parameters of the dimension. This represents the maximum sum of the absolute values ​​of all dimension parameters. , This represents the total dimension of the original parameters. This sampling rate can increase the sampling ratio for sparse dimensions, adapting to the characteristics of high-dimensional heterogeneous parameters. The value range is [0, 1]. The smaller, The smaller the value, the higher the sampling ratio, in order to improve the anomaly detection accuracy of sparse dimensions.

[0044] In this embodiment, the preset abnormal threshold can be set according to needs or experience, for example, it can take values ​​in the range of [0.7, 0.9], where the value for high-risk scenarios is 0.9. When the first When the abnormal score of a dimension parameter exceeds the preset abnormal threshold, the dimension will be marked as an abnormal dimension and deleted. The remaining set of valid dimensions will be used for subsequent filtering and verification.

[0045] Optionally, to reduce the interference of extreme values ​​on parameter verification, the verification node performs parameter standardization and robustness transformation on the remaining valid dimension parameters.

[0046] Parameter standardization and robustness are transformed into parameters of the remaining effective dimensions after pre-screening for abnormal parameters, and standardized using the median and median absolute deviation (MAD). The formula for standardization is: ; in, Indicates the first The first organization uploaded The standardized parameters of the dimension parameters, after standardization, The mean is close to 0, and the variance is close to 1, eliminating the influence of dimensions and extreme values. Indicates the first The first organization uploaded Dimensional parameters, Indicates the first The first organization uploaded The median of the parameter set of the dimension. Indicates the first The first organization uploaded The median absolute deviation of the parameters in a dimension is used to measure the degree of parameter dispersion and has extremely strong noise resistance.

[0047] Optionally, to avoid the singularity problem of high-dimensional parameter matrices and improve the stability and computational efficiency of parameter distribution consistency testing, the verification node adopts an improved Mahalanobis distance algorithm that introduces regularized covariance estimation and feature selection mechanisms to calculate the consistency between the standardized parameter vector and the global parameters.

[0048] The multidimensional consistency test is an improved Mahalanobis distance that introduces regularized covariance estimation and feature selection mechanisms. It calculates the Mahalanobis distance between the standardized parameters and the global parameters to ensure consistency between individual parameters and global parameters. When the obtained Mahalanobis distance is greater than a preset dynamic threshold, the current standardized parameter is identified as an outlier and marked. The preset dynamic threshold is a dynamic threshold set based on the chi-square distribution. The specific method for determining the preset dynamic threshold is as follows: based on the chi-square distribution, the chi-square quantile corresponding to the 95% confidence level is taken as the preset dynamic threshold, which can be obtained by querying the chi-square distribution table or by calling statistical tools.

[0049] The improved formula for calculating Mahalanobis distance is: ; in, Indicates the first The Mahalanobis distance between the standardized parameters of an individual institution and the global parameters indicates that the larger the value, the worse the consistency of parameter distribution and the higher the probability of anomalies. Indicates the first Standardized parameters for each institution, with dimensions of , This represents the valid dimensions after filtering in step 101. This represents the global parameter mean vector, with dimension . It is composed of the mean values ​​of each dimension of the standardized parameters of all institutions, the first... Dimensional elements , This indicates the first institution to upload the first [item / file]. The mean of the parameter set of the dimension. This indicates that the second organization uploaded the first... The mean of the parameter set of the dimension. Indicates the first The first organization uploaded Using the median instead of the mean of the parameter set for a given dimension can further improve noise robustness. (This is followed by a seemingly unrelated sentence about traversing the data.) Construct all dimensions ; This represents the global parameter covariance matrix after feature filtering. This represents the regularization coefficient, which can range from [0.01, 0.1]. It is used to avoid singularity in the covariance matrix and improve matrix invertibility. A value of 0.1 is used when the covariance matrix fluctuates greatly, and 0.01 is used when the fluctuation is small. This represents the identity matrix, where the diagonal elements are all 1s and all other elements are 0s. This represents the inverse of the regularized covariance matrix.

[0050] Optionally, to adapt to the non-uniformity of parameter spatial distribution, the verification node adopts the optimized Local Outlier Factor (LOF) algorithm, which introduces dynamic neighborhood selection and weighted distance measurement mechanism, to perform fine-grained anomaly detection at the local neighborhood level for parameters.

[0051] Fine-grained detection employs an optimized local anomaly factor algorithm that incorporates dynamic neighborhood selection and weighted distance measurement mechanisms. It calculates the local anomaly factor for the standardized parameters after multidimensional consistency testing, and determines that the current standardized parameter is anomaly in the local neighborhood when the local anomaly factor is greater than a preset anomaly factor threshold. The formula for calculating the local anomaly factor optimization algorithm is as follows: ; in, Indicates the first The optimized local anomaly factor for the standardized parameters of each institution indicates that the parameters are abnormal when the value is greater than 1, and the higher the value, the greater the degree of anomaly. Indicates the first Standardized parameters for each organization nearest neighbor set Indicates the first Standardized parameters for each organization The number of elements in the nearest neighbor set. Indicates distance weight, Represents the nearest neighbor set of the i-th Standardized parameter vectors for each institution, This represents the neighborhood scaling parameter, which can take any multiple from 0.5 to 1 times the global standard deviation of the parameter. It is used to assign distance decay weights to nearest neighbors. Represents the nearest neighbor set of the i-th The local reachability density of an individual mechanism reflects the parameter The parameter density in the region; the lower the density, the smaller the value. Indicates the first The local reachability density of standardized parameters for an organization.

[0052] The preset abnormal factor threshold can be flexibly adjusted according to the parameter distribution characteristics. In this embodiment, its value is not limited. For example, the preset abnormal factor threshold can be 1.5, 1.6, etc.

[0053] After completing the four-stage parameter validity screening process, each verification node uploads its initial verification results to the smart contract. Once the smart contract collects the initial verification results from all verification nodes, it initiates a dynamic voting mechanism based on a dual weighting of similarity and contribution to aggregate the results from all verification nodes, forming a final consensus. This resolves the subjectivity and misjudgment issues inherent in single-node verification, improving the reliability and attack resistance of the consensus result.

[0054] In one embodiment, the final consensus verification result is determined through a multi-node consensus mechanism, including: according to Determine the final consensus verification result; in, Indicates the first The final consensus verification results of the standardized parameters of each institution. This is a symbolic function used to output a clear decision result after consensus (e.g., pass or fail). When the value within the parentheses is greater than 0, then... A value of 1 indicates that the final consensus verification result is passed and the model can participate in global model aggregation. A value of 0 within the parentheses indicates that... A value of 0 indicates that re-verification is required. If the value within the parentheses is less than 0, then... A value of -1 indicates that the final consensus verification result is unsuccessful, and the system refuses to participate in the global model aggregation. Indicates the first The similarity weight of each verification node reflects the degree of consistency between the initial verification result of that node and the results of other nodes. , This represents the total number of verification nodes. This represents the similarity enhancement coefficient, which takes a value greater than 1. Optionally, when its value is 2, it can effectively enhance the weight of nodes with high consensus consistency, while avoiding excessive concentration of weight. Indicates the first The consistency between the verification results of a verification node and the verification results of other verification nodes can be calculated using improved cosine similarity. Indicates the first The historical verification credibility coefficient of each verification node; Indicates the first The verification node verifies the first... Initial validation results of standardized parameters for each organization. The values ​​are -1, 0, and 1 respectively. A value of -1 indicates that the initial verification failed, a value of 1 indicates that the initial verification passed, and a value of 0 indicates that the initial verification is pending. This represents the contribution adjustment coefficient. Indicates the first The contribution weight of each verification node. It equals the quotient of the historical contribution divided by the maximum contribution in the contribution set, where the historical contribution equals the total number of times the verification node participated in verification multiplied by the average verification accuracy and then multiplied by the average computing power contribution.

[0055] Through the aforementioned dual-weighted dynamic voting mechanism, the smart contract ultimately determines the final consensus verification result of the high-dimensional heterogeneous parameters of each local model, ensuring that the final consensus verification result can comprehensively reflect the verification capabilities and historical performance of each verification node, and has both accuracy and fairness.

[0056] Step 104: The smart contract writes the final consensus verification result, parameter details and verification records of each node into the blockchain for storage. Parameters that pass the authorization verification participate in the global model aggregation, and parameters that fail the verification are rejected.

[0057] All the aforementioned verification algorithms and mathematical models are embedded in a blockchain smart contract. Once the verification process is triggered, the smart contract automatically completes the selection of verification nodes, parameter standardization, consistency verification, and consensus voting among multiple nodes. The final consensus verification result, parameter details, and voting records of each node are written into the blockchain block, forming an immutable verification record. Parameters that pass verification are authorized to participate in global model aggregation; parameters that fail verification are rejected, triggering anomaly alerts and a node trust level downgrade mechanism.

[0058] Optionally, the information that smart contracts automatically write to the blockchain includes: Parameter details: High-dimensional heterogeneous parameters of the original local model Effective Dimension Set Standardized parameters ; Verification node information: Set of verification nodes Trust level of each node Similarity weight Contribution weight ; Verification process data: Anomaly score Mahalanobis distance Initial verification results ; Final result: Voting statistics details (number of nodes that passed / failed / pending and their weight percentage).

[0059] All information is stored according to the blockchain block structure and is encrypted with hashes to ensure immutability. Subsequent audits can query the complete verification process by block height or transaction ID to achieve accountability.

[0060] This invention provides a method for verifying high-dimensional heterogeneous parameters of a federated learning model. A smart contract selects institutions meeting trust requirements from among the participating institutions in a multi-centralized federated learning system as verification nodes, forming a verification node set. After each institution completes local model training, it uploads the high-dimensional heterogeneous parameters of that local model to the blockchain, triggering the smart contract verification process. Verification nodes verify all uploaded high-dimensional heterogeneous parameters of the local model. The smart contract collects the initial verification results from each verification node and determines the final consensus verification result through a multi-node consensus mechanism. The smart contract writes the final consensus verification result, parameter details, and verification records of each node into the blockchain for notarization. Parameters that pass verification are authorized to participate in global model aggregation, while parameters that fail verification are rejected. This invention, by selecting institutions meeting trust requirements as verification nodes through smart contracts for subsequent verification of high-dimensional heterogeneous parameters of the local model, replaces the traditional third-party parameter aggregation model, completely resolving the risk of single-point failures and avoiding global model pollution caused by dishonest third-party behavior, thereby improving the stability and security of the federated learning system. This invention employs a quantitative trust screening system based on historical verification accuracy, computing power contribution, and compliance scores to ensure high credibility of participating nodes. Combined with a multi-node consensus mechanism to aggregate verification results, it effectively resists malicious attacks such as data discrimination and model poisoning, guaranteeing the authenticity and validity of local model parameters. This invention uses smart contracts to completely write the final consensus verification results, parameter details, and verification records of each node into the blockchain for notarization, forming a full-process audit chain. This solves the problems of difficult traceability and ambiguous responsibility definition in traditional verification methods, providing a reliable compliance basis for cross-institutional collaboration. This invention's verification process is designed for high-dimensional heterogeneous parameter characteristics. Through multi-stage verification logic and a consensus mechanism, it balances verification accuracy and efficiency, ensuring that in cross-domain collaboration scenarios such as finance and healthcare, the verification effect of complex parameters is unaffected by dimensionality and heterogeneity, guaranteeing the quality of global model aggregation.

[0061] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0062] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0063] Figure 2 The diagram shows a schematic of a verification device for high-dimensional heterogeneous parameters of a federated learning model according to an embodiment of the present invention. For ease of explanation, only the parts relevant to the embodiment of the present invention are shown, and are described in detail below: like Figure 2As shown, the verification device for high-dimensional heterogeneous parameters of the federated learning model includes: smart contract module 21, institution terminal 22, and verification node terminal 23.

[0064] Smart contract module 21 is used to select institutions that meet the trust requirements from among the institutions participating in multi-centralized federated learning as verification nodes, forming a set of verification nodes. Each institution's terminal 22 is used to upload the high-dimensional heterogeneous parameters of the local model to the blockchain after completing the local model training, triggering the smart contract verification process; Verification node terminal 23 is used to verify all the high-dimensional heterogeneous parameters of the uploaded local model and send the initial verification results to smart contract module 21; The smart contract module 21 is also used to collect the initial verification results of each verification node, determine the final consensus verification result through a multi-node consensus mechanism, and write the final consensus verification result, parameter details and verification records of each node into the blockchain for storage. It also authorizes parameters that have passed verification to participate in the global model aggregation and rejects parameters that have not passed verification.

[0065] In one possible implementation, when the smart contract module 21 selects an institution that meets the trust requirements as a verification node from among the institutions participating in the multi-centralized federated learning, it is used for: The smart contract obtains the historical verification accuracy, computing power contribution value, and compliance score of each institution participating in the multi-centralized federated learning; An improved PBFT algorithm is used to calculate the trust score of each institution based on historical verification accuracy, computing power contribution value, and compliance score. When the trust value is greater than or equal to the trust threshold, the institution corresponding to the trust value is determined as the verification node.

[0066] In one possible implementation, the smart contract module 21 employs a modified PBFT algorithm to calculate the trust value of each institution based on historical verification accuracy, computing power contribution value, and compliance score, for the following purposes: according to Calculate the trust score for each institution; in, Indicates the first Trust rating of an organization , , These are the weighting coefficients, and , Indicates the first Historical verification accuracy of individual institutions Indicates the first The computing power contribution of each institution. Indicates the first Compliance scores for each organization.

[0067] In one possible implementation, when the verification node terminal 23 verifies all uploaded local model high-dimensional heterogeneous parameters, it is used for: For all uploaded local model high-dimensional heterogeneous parameters, perform parameter anomaly dimension pre-screening, parameter standardization and robust transformation, multidimensional consistency test and fine-grained detection in sequence.

[0068] In one possible implementation, the parameter anomaly dimension pre-screening is performed by using an improved parallel isolated forest algorithm to calculate anomaly scores for each dimension of the high-dimensional heterogeneous parameters of the local model. When the anomaly score is greater than a preset anomaly threshold, the current dimension is marked as an anomaly dimension and deleted. The improved parallel isolated forest algorithm introduces time decay weights and a sparse adaptive sampling mechanism on the basis of the existing isolated forest algorithm. The calculation formula for the improved parallel isolated forest algorithm is as follows: ; in, Indicates the first Outlier scores in dimensions Indicates the first The first organization uploaded The set of parameters for the dimension. Indicates the time decay weight. , This represents the difference between the timestamp of the current parameter upload and the system's base time. Indicates the preset time decay period. Indicates a sparse adaptive sampling rate. , Indicates the first The first organization uploaded The sum of the absolute values ​​of all parameters of the dimension. This represents the maximum value of the sum of the absolute values ​​of all dimension parameters.

[0069] In one possible implementation, parameter standardization and robustness are transformed into standardizing the parameters of the remaining dimensions after pre-screening for outlier parameters using the median and the absolute deviation of the median. The formula for standardization is: ; in, Indicates the first The first organization uploaded Standardized parameters of dimensional parameters, Indicates the first The first organization uploaded Dimensional parameters, Indicates the first The first organization uploaded The median of the parameter set of the dimension. Indicates the first The first organization uploaded The median absolute deviation of the parameters of the dimension.

[0070] In one possible implementation, the multidimensional consistency test is to use an improved Mahalanobis distance that introduces regularized covariance estimation and feature selection mechanism to calculate the Mahalanobis distance between the standardized parameter and the global parameter, and when the Mahalanobis distance is greater than a preset dynamic threshold, the current standardized parameter is determined to be an outlier and marked. The improved formula for calculating Mahalanobis distance is: ; in, Indicates the first Mahalanobis distance between the standardized parameters of each institution and the global parameters. Indicates the first Standardized parameters for each organization This represents the global parameter mean vector. This represents the global parameter covariance matrix after feature filtering. Represents the regularization coefficient. Represents the identity matrix.

[0071] In one possible implementation, fine-grained detection employs an optimized local anomaly factor algorithm that incorporates dynamic neighborhood selection and weighted distance measurement mechanisms. This algorithm calculates the local anomaly factor for the standardized parameters after multidimensional consistency testing, and determines that the current standardized parameter is anomaly in the local neighborhood when the local anomaly factor exceeds a preset anomaly factor threshold. The formula for calculating the local anomaly factor optimization algorithm is as follows: ; in, Indicates the first Local outlier factors after optimization of standardized parameters for each institution. Indicates the first Standardized parameters for each organization nearest neighbor set Indicates the first Standardized parameters for each organization The number of elements in the nearest neighbor set. Indicates distance weight, Represents the nearest neighbor set of the i-th Standardized parameter vectors for each institution, Represents the neighborhood scale parameter. Represents the nearest neighbor set of the i-th The locally achievable density of an individual institution. Indicates the first The local reachability density of standardized parameters for an organization.

[0072] In one possible implementation, when the smart contract module 21 determines the final consensus verification result through a multi-node consensus mechanism, it is used for: according to Determine the final consensus verification result; in, Indicates the first The final consensus verification results of the standardized parameters of each institution. Indicates the first The similarity weight of each verification node. , This represents the total number of verification nodes. Represents the similarity enhancement coefficient. Indicates the first The consistency of the verification results of each verification node with the verification results of other verification nodes. This represents the historical verification reliability coefficient of the verification node. Indicates the first The verification node verifies the first... Initial validation results of standardized parameters for each organization. This represents the contribution adjustment coefficient. Indicates the first The contribution weight of each verification node.

[0073] The above embodiments provide a verification device for high-dimensional heterogeneous parameters of a federated learning model. A smart contract module selects institutions meeting trust requirements from among the participating institutions in a multi-centralized federated learning model as verification nodes, forming a verification node set. After each institution's terminal completes local model training, it uploads the high-dimensional heterogeneous parameters of the local model to the blockchain, triggering the smart contract verification process. The verification node terminals verify all uploaded high-dimensional heterogeneous parameters of the local model. The smart contract module collects the initial verification results from each verification node and determines the final consensus verification result through a multi-node consensus mechanism. The final consensus verification result, parameter details, and verification records of each node are written to the blockchain for notarization. Parameters that pass verification are authorized to participate in global model aggregation, while parameters that fail are rejected. This embodiment of the invention ultimately selects institutions meeting trust requirements as verification nodes through smart contracts for subsequent verification of high-dimensional heterogeneous parameters of the local model, replacing the traditional third-party parameter aggregation mode. This completely solves the potential for single-point failures and avoids global model pollution caused by dishonest third-party behavior, improving the stability and security of the federated learning system. This invention employs a quantitative trust screening system based on historical verification accuracy, computing power contribution, and compliance scores to ensure high credibility of participating nodes. Combined with a multi-node consensus mechanism to aggregate verification results, it effectively resists malicious attacks such as data discrimination and model poisoning, guaranteeing the authenticity and validity of local model parameters. This invention uses smart contracts to completely write the final consensus verification results, parameter details, and verification records of each node into the blockchain for notarization, forming a full-process audit chain. This solves the problems of difficult traceability and ambiguous responsibility definition in traditional verification methods, providing a reliable compliance basis for cross-institutional collaboration. This invention's verification process is designed for high-dimensional heterogeneous parameter characteristics. Through multi-stage verification logic and a consensus mechanism, it balances verification accuracy and efficiency, ensuring that in cross-domain collaboration scenarios such as finance and healthcare, the verification effect of complex parameters is unaffected by dimensionality and heterogeneity, guaranteeing the quality of global model aggregation.

[0074] Figure 3 This is a schematic diagram of a terminal provided in an embodiment of the present invention. Figure 3 As shown, the terminal 3 in this embodiment includes a processor 30, a memory 31, and a computer program 32 stored in the memory 31 and executable on the processor 30. When the processor 30 executes the computer program 32, it implements the steps in the embodiments of the verification methods for high-dimensional heterogeneous parameters of the various federated learning models described above, for example... Figure 1 Steps 101 to 104 are shown. Alternatively, when processor 30 executes computer program 32, it implements the functions of each module / unit in the above-described device embodiments, for example... Figure 2 The functions of each module / unit are shown.

[0075] For example, computer program 32 can be divided into one or more modules / units, one or more of which are stored in memory 31 and executed by processor 30 to complete the present invention. One or more modules / units can be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of computer program 32 in terminal 3. For example, computer program 32 can be divided into... Figure 2 The modules / units shown are shown.

[0076] Terminal 3 may include, but is not limited to, processor 30 and memory 31. Those skilled in the art will understand that... Figure 3 This is merely an example of terminal 3 and does not constitute a limitation on terminal 3. It may include more or fewer components than shown, or combine certain components, or different components. For example, the terminal may also include input / output devices, network access devices, buses, etc.

[0077] The processor 30 may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0078] The memory 31 can be an internal storage unit of the terminal 3, such as a hard disk or RAM of the terminal 3. The memory 31 can also be an external storage device of the terminal 3, such as a plug-in hard disk, Smart MediaCard (SMC), Secure Digital (SD) card, or Flash Card equipped on the terminal 3. Furthermore, the memory 31 can include both internal and external storage units of the terminal 3. The memory 31 is used to store computer programs and other programs and data required by the terminal. The memory 31 can also be used to temporarily store data that has been output or will be output.

[0079] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0080] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0081] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0082] In the embodiments provided by this invention, it should be understood that the disclosed devices / terminals and methods can be implemented in other ways. For example, the device / terminal embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0083] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0084] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0085] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of the above embodiments of the verification methods for high-dimensional heterogeneous parameters of various federated learning models. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0086] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for validating high-dimensional heterogeneous parameters of a federated learning model, characterized in that, include: The smart contract selects institutions that meet the trust requirements from among the participating institutions in the multi-centralized federated learning process to form a set of verification nodes. After each institution completes the training of its local model, it uploads the high-dimensional heterogeneous parameters of the local model to the blockchain, triggering the smart contract verification process. The verification node verifies all the high-dimensional heterogeneous parameters of the uploaded local model. The smart contract collects the initial verification results of each verification node and determines the final consensus verification result through a multi-node consensus mechanism. The smart contract writes the final consensus verification result, parameter details, and verification records of each node into the blockchain for storage. Parameters that pass the verification are authorized to participate in the global model aggregation, while parameters that fail the verification are rejected.

2. The method for verifying high-dimensional heterogeneous parameters of a federated learning model according to claim 1, characterized in that, The smart contract selects institutions that meet the trust requirements from among the participating institutions in the multi-centralized federated learning process as verification nodes, including: The smart contract obtains the historical verification accuracy, computing power contribution value, and compliance score of each institution participating in the multi-centralized federated learning; An improved PBFT algorithm is used to calculate the trust score of each institution based on the historical verification accuracy, computing power contribution value, and compliance score. When the trust value is greater than or equal to the trust threshold, the institution corresponding to the trust value is determined as a verification node.

3. The method for verifying high-dimensional heterogeneous parameters of a federated learning model according to claim 2, characterized in that, An improved PBFT algorithm is used to calculate the trust score of each institution based on the historical verification accuracy, computing power contribution value, and compliance score, including: according to Calculate the trust score for each institution; in, Indicates the first Trust rating of an organization , , These are the weighting coefficients, and , Indicates the first Historical verification accuracy of individual institutions Indicates the first The computing power contribution of each institution. Indicates the first Compliance scores for each organization.

4. The method for verifying high-dimensional heterogeneous parameters of a federated learning model according to any one of claims 1-3, characterized in that, The verification node verifies all uploaded local model high-dimensional heterogeneous parameters, including: The verification node sequentially performs parameter anomaly dimension pre-screening, parameter standardization and robust conversion, multidimensional consistency check and fine-grained detection on all uploaded local model high-dimensional heterogeneous parameters.

5. The method for verifying high-dimensional heterogeneous parameters of a federated learning model according to claim 4, characterized in that, The parameter anomaly dimension pre-screening is performed by using an improved parallel isolated forest algorithm to calculate anomaly scores for each dimension of the high-dimensional heterogeneous parameters of the local model. When the anomaly score is greater than a preset anomaly threshold, the current dimension is marked as an anomaly dimension and deleted. The improved parallel isolated forest algorithm introduces time decay weights and a sparse adaptive sampling mechanism on the basis of the existing isolated forest algorithm. The calculation formula for the improved parallel isolated forest algorithm is as follows: ； in, Indicates the first Outlier scores in dimensions Indicates the first The first organization uploaded The set of parameters for the dimension. Indicates the time decay weight. , This represents the difference between the timestamp of the current parameter upload and the system's base time. Indicates the preset time decay period. Indicates a sparse adaptive sampling rate. , Indicates the first The first organization uploaded The sum of the absolute values ​​of all parameters of the dimension. This represents the maximum value of the sum of the absolute values ​​of all dimension parameters.

6. The method for verifying high-dimensional heterogeneous parameters of a federated learning model according to claim 5, characterized in that, The parameter standardization and robustness are converted into parameters of the remaining dimensions after the parameter outlier dimension pre-screening, and standardized using the median and median absolute deviation. The formula for standardization is: ； in, Indicates the first The first organization uploaded Standardized parameters of dimensional parameters, Indicates the first The first organization uploaded Dimensional parameters, Indicates the first The first organization uploaded The median of the parameter set of the dimension. Indicates the first The first organization uploaded The median absolute deviation of the parameters of the dimension.

7. The method for verifying high-dimensional heterogeneous parameters of a federated learning model according to claim 6, characterized in that, The multidimensional consistency test employs an improved Mahalanobis distance with regularized covariance estimation and feature selection mechanisms to calculate the Mahalanobis distance between the standardized parameters and the global parameters. When the Mahalanobis distance is greater than a preset dynamic threshold, the current standardized parameter is determined to be an abnormal parameter and marked. The improved formula for calculating Mahalanobis distance is: ； in, Indicates the first Mahalanobis distance between the standardized parameters of each institution and the global parameters. Indicates the first Standardized parameters for each organization This represents the global parameter mean vector. This represents the global parameter covariance matrix after feature filtering. Represents the regularization coefficient. Represents the identity matrix.

8. The method for verifying high-dimensional heterogeneous parameters of a federated learning model according to claim 7, characterized in that, The fine-grained detection employs an optimized local anomaly factor algorithm that incorporates dynamic neighborhood selection and weighted distance measurement mechanisms. This algorithm calculates the local anomaly factor for the standardized parameters after multidimensional consistency testing. When the local anomaly factor exceeds a preset anomaly factor threshold, it determines that the current standardized parameter is anomaly in the local neighborhood. The calculation formula for the optimized local anomaly factor algorithm is as follows: ； in, Indicates the first Local outlier factors after optimization of standardized parameters for each institution. Indicates the first Standardized parameters for each organization Nearest neighbor set Indicates the first Standardized parameters for each organization The number of elements in the nearest neighbor set. Indicates distance weight, Represents the nearest neighbor set of the i-th Standardized parameter vectors of each institution Represents the neighborhood scale parameter. Represents the nearest neighbor set of the i-th The locally achievable density of an individual institution. Indicates the first The local reachability density of standardized parameters for an organization.

9. The method for verifying high-dimensional heterogeneous parameters of a federated learning model according to claim 8, characterized in that, The final consensus verification result is determined through a multi-node consensus mechanism, including: according to Determine the final consensus verification result; in, Indicates the first The final consensus verification results of the standardized parameters of each institution. Indicates the first The similarity weight of each verification node. , This represents the total number of verification nodes. Represents the similarity enhancement coefficient. Indicates the first The consistency of the verification results of each verification node with the verification results of other verification nodes. This represents the historical verification reliability coefficient of the verification node. Indicates the first The verification node verifies the first... Initial validation results of standardized parameters for each organization. This represents the contribution adjustment coefficient. Indicates the first The contribution weight of each verification node.

10. A verification device for high-dimensional heterogeneous parameters of a federated learning model, characterized in that, include: The smart contract module is used to select institutions that meet the trust requirements from among the participating institutions in the multi-centralized federated learning as verification nodes, forming a set of verification nodes. Each institution's terminal is used to upload the high-dimensional heterogeneous parameters of the local model to the blockchain after completing the local model training, triggering the smart contract verification process; The verification node terminal is used to verify the high-dimensional heterogeneous parameters of all uploaded local models and send the initial verification results to the smart contract module. The smart contract module is also used to collect the initial verification results of each verification node, determine the final consensus verification result through a multi-node consensus mechanism, and write the final consensus verification result, parameter details and verification records of each node into the blockchain for storage. Parameters that pass the verification are authorized to participate in the global model aggregation, while parameters that fail the verification are rejected.

Images